A sleep center information integrated digital diagnosis and treatment system

Abstract

The specification provides a sleep center information integrated digital diagnosis and treatment system, the system is configured with a sleep disorder database and a sleep diagnosis and treatment large model, the sleep disease database stores the electronic medical record information data of the patient, the original data collected by the medical equipment and the sleep disorder data collected by the terminal equipment, the sleep diagnosis and treatment large model recommends the diagnosis and treatment result based on the data stored in the sleep disease database, the system comprises: a determination module for determining the current sleep disorder diagnosis and treatment recommendation corresponding to the diagnosis and treatment task; a calling module for calling the electronic medical record information data corresponding to the current patient stored in the sleep disorder database, the image data collected by the medical equipment and the sleep monitoring data collected by the terminal equipment, and generating the corresponding image analysis report and sleep evaluation report; a recommendation module for inputting the sleep evaluation report, the image analysis report and the corresponding electronic medical record information data of the patient into the sleep diagnosis and treatment large model to obtain the diagnosis and treatment recommendation result corresponding to the diagnosis and treatment task.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of computer technology, and in particular to an integrated digital diagnosis and treatment system for sleep center information. Background Technology

[0002] With the fast pace and increasing pressure of modern life, sleep disorders have become a prevalent health problem. To better diagnose and treat sleep disorders, many medical institutions have established specialized sleep centers, providing a range of services from monitoring to treatment. However, existing sleep centers suffer from several shortcomings in data management and information integration, primarily in the following aspects: Data silos: Sleep centers typically include multiple departments, such as monitoring control rooms, research information control rooms, non-drug treatment rooms, traditional Chinese medicine treatment rooms, testing rooms, wards, and reception rooms. These departments use different information systems, leading to ineffective data integration and sharing. Cumbersome manual operation: Doctors and staff need to switch between different systems, manually entering and synchronizing data, which not only increases workload but also increases the risk of errors. Incomplete information: Due to the lack of a unified data management platform, patient information may be scattered across various systems, making it difficult to form complete medical records. Low diagnostic and treatment efficiency: Doctors cannot quickly obtain comprehensive patient information, affecting the speed and accuracy of diagnostic and treatment decisions. Summary of the Invention

[0003] To overcome the problems existing in related technologies, this specification provides an integrated digital diagnosis and treatment system for sleep center information.

[0004] According to a first aspect of the embodiments of this specification, an integrated digital diagnosis and treatment system for sleep centers is provided. The system is configured with a sleep disorder specialty database and a large-scale sleep diagnosis and treatment model. The sleep disorder database stores patients' electronic medical record information, raw data collected by medical devices, and sleep disorder data collected by terminal devices. The large-scale sleep diagnosis and treatment model recommends treatment results based on the data stored in the sleep disorder database. The system includes: The determination module is used to determine the treatment task corresponding to the current sleep disorder treatment recommendation; The calling module is used to call the electronic medical record information data stored in the sleep disorder disease database corresponding to the current patient, the image data collected by medical devices, and the sleep monitoring data collected by terminal devices, and generate corresponding image analysis reports and sleep assessment reports. The recommendation module is used to input the patient's sleep assessment report, image analysis report and corresponding electronic medical record information into the sleep diagnosis and treatment model to obtain the diagnosis and treatment recommendation results corresponding to the diagnosis and treatment task.

[0005] Based on the above embodiments in this specification, it can be seen that: This manual describes an integrated digital diagnosis and treatment system for sleep centers. Through a configured sleep disorder-specific database, it integrates data from different departments onto a unified digital diagnosis and treatment platform, achieving centralized data management and sharing. An automated data synchronization mechanism reduces manual operation and improves data accuracy and timeliness. It establishes complete electronic patient records, including medical history, examination results, treatment plans, and all other relevant information. Utilizing big data and artificial intelligence technologies, based on a large-scale sleep diagnosis and treatment model, it analyzes patient sleep data and provides personalized treatment recommendations. It also provides an easy-to-use user interface for convenient operation and viewing of relevant information by doctors and patients.

[0006] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this specification. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this specification and, together with the description, serve to explain the principles of this specification.

[0008] Figure 1 This is a system architecture diagram of an integrated digital diagnosis and treatment system for sleep center information provided in an exemplary embodiment of this specification.

[0009] Figure 2 This is a block diagram of an integrated digital diagnosis and treatment system for sleep center information provided in an exemplary embodiment of this specification.

[0010] Figure 3 This is a flowchart illustrating a method for establishing a sleep disorder-specific database, provided in an exemplary embodiment of this specification.

[0011] Figure 4 This is a flowchart illustrating text processing based on a knowledge graph embedding component, provided in an exemplary embodiment of this specification.

[0012] Figure 5 This is a schematic diagram of an electronic device provided in an exemplary embodiment of this specification. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification.

[0014] It should be noted that in other embodiments, the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification. In some other embodiments, the methods may include more or fewer steps than those described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments. It should be understood that although the terms first, second, third, etc., may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this specification, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0015] Figure 1 This specification provides a system architecture diagram of an integrated digital diagnosis and treatment system for sleep centers, as illustrated in an exemplary embodiment. The system may include a terminal 10, a server 11 configured with the integrated digital diagnosis and treatment system for sleep centers, and a sleep disorder database 12. The system may be configured with a sleep disorder database and a comprehensive sleep diagnosis and treatment model. The sleep disorder database stores patients' electronic medical record information, raw data collected by medical devices, and sleep disorder data collected by the terminal devices. The comprehensive sleep diagnosis and treatment model recommends treatment outcomes based on the data stored in the sleep disorder database.

[0016] Terminal 10 can access the integrated digital diagnosis and treatment system for sleep centers described in this manual, thereby selecting and executing tasks through terminal control. Terminal 10 can also be a mobile phone, smartwatch, VR device, etc., and this manual does not impose any limitations on it.

[0017] Server 11 can be a physical server containing a standalone host, or it can be a virtual server (such as a cloud server) hosted in a host cluster. Server 11 can run the integrated digital diagnosis and treatment system for sleep centers described in this specification. Through a configured sleep disorder specialty database, it can integrate data from different departments onto a unified integrated digital diagnosis and treatment system platform, achieving centralized data management and sharing. An automated data synchronization mechanism reduces manual operations and improves data accuracy and timeliness. Complete electronic patient records are established, including the patient's medical history, examination results, treatment plans, and all other relevant information. Utilizing big data and artificial intelligence technologies, based on a large-scale sleep diagnosis and treatment model, it analyzes patient sleep data and provides personalized treatment suggestions. It also provides an easy-to-use user interface for doctors and patients to operate and view relevant information.

[0018] Figure 2 This is a block diagram of an integrated digital diagnosis and treatment system for sleep centers, provided in an exemplary embodiment of this specification. The system is configured with a sleep disorder specialty database and a large-scale sleep diagnosis and treatment model. The sleep disorder database stores patients' electronic medical record information, raw data collected by medical devices, and sleep disorder data collected by terminal devices. The large-scale sleep diagnosis and treatment model recommends treatment outcomes based on the data stored in the sleep disorder database, and may specifically include the following modules: The determination module 202 is used to determine the treatment task corresponding to the current sleep disorder treatment recommendation; The diagnostic and treatment tasks described in this manual may include at least one of the following: diagnostic reference, treatment recommendations, lifestyle modification recommendations, complication risk assessment, treatment effectiveness evaluation, and / or clinical pathway development. For example, if a patient needs diagnostic reference, they can select "diagnostic reference" as the current diagnostic and treatment task. Of course, multiple diagnostic and treatment tasks can be selected; for example, both "diagnostic reference" and "treatment recommendations" can be selected as the current diagnostic and treatment tasks simultaneously. This manual does not impose any restrictions on this.

[0019] The diagnostic and treatment tasks described in this specification can be selected by the patient or doctor through a task list, thereby reducing noise and errors that may be introduced by free text input, improving the stability and accuracy of the diagnostic and treatment system. Focusing on a limited range of tasks can significantly improve the accuracy of fine-tuning large models, making them more professional and reliable in specific diagnostic and treatment areas, and improving the model's response speed and processing efficiency. At the same time, the list selection method simplifies the user operation process, making it more convenient for doctors to use the system. Of course, in some embodiments, the diagnostic and treatment tasks described in this specification can also be manually entered by the user.

[0020] The calling module 204 is used to call the electronic medical record information data stored in the sleep disorder disease database corresponding to the current patient, the image data collected by the medical device, and the sleep monitoring data collected by the terminal device, and generate corresponding image analysis reports and sleep assessment reports.

[0021] The sleep disorder database stores patients' electronic medical records, image data collected by medical devices, and sleep monitoring data collected by terminal devices. For ease of understanding, this manual describes how to establish the sleep disorder database, specifically including the following steps: Step 302: Obtain the patient's electronic medical record information data and the raw data collected by the medical device.

[0022] The sleep disorder database described in this manual has at least one storage node deployed.

[0023] The electronic medical record information can include the patient's basic information (height and weight), medical history, follow-up records, treatment records, etc. Medical equipment can refer to the medical diagnostic equipment (such as PSG, X-ray diagnostic equipment, ultrasound diagnostic equipment, etc.), medical treatment equipment (such as ventilators, hyperbaric oxygen chambers, etc.), and auxiliary equipment (such as central suction and oxygen supply systems, blood bank equipment, etc.) configured in the sleep center. During use, these devices can collect raw data corresponding to the patient or user. For example, taking PSG as an example, it can collect the patient's electroencephalogram (EEG) data in real time.

[0024] It is understandable that the aforementioned electronic medical record information data may also include data filled in by the patient. For example, when a user experiences (or uses) a medical device, the patient can scan the QR code corresponding to the device to fill in information. This information can also be used as the patient's electronic medical record information and obtained together.

[0025] Step 304: Obtain sleep disorder data after edge computing collected by the terminal device corresponding to the patient.

[0026] Terminal devices can collect patients' sleep disorder data. For example, the wristband 12 can collect the patient's heart rate, pulse, body temperature, and blood oxygen concentration, which can be used as the sleep disorder data described in this manual. Similarly, the mobile phone 10 can collect the patient's breathing sounds, such as snoring, and process them into corresponding audio data, which can also be used as the sleep disorder data described in this manual. Furthermore, the VR headset 11 can collect the patient's eye movement data, which can also be used as the sleep disorder data described in this manual.

[0027] It is understandable that the sleep disorder data collected by the patient terminal device may be abnormal, such as data loss, data saturation, or high interference noise due to long sleep data collection time. Therefore, in order to ensure accuracy, in one embodiment, the sleep disorder disease database can be configured with an anomaly detection component. The anomaly detection component can be used to determine whether the sleep disorder data is abnormal, so as to remove the abnormal sleep disorder data and thus improve the accuracy of the data.

[0028] The patient's sleep environment can also interfere with the sleep disorder data. Therefore, when collecting data, the terminal device can simultaneously record environmental data. For example, the terminal device can additionally record data such as temperature and sound in the patient's environment. The above environmental data can be transmitted back to the sleep disorder database, which then uses an anomaly detection component to detect and eliminate interference from the sleep environment.

[0029] Step 306: Based on the configured first and second network cards, perform isolation and secure communication of the data inside and outside the hospital. The first network card is used to connect to the intranet corresponding to the sleep center to exchange data within the hospital, and the second network card is used to connect to the corresponding extranet to exchange data outside the hospital.

[0030] The terminal device can transmit collected sleep disorder data (such as heart rate, blood oxygen, and EEG) to the second network card corresponding to the sleep disorder database via wireless communication protocols. Specifically, the terminal device can communicate with the sleep disorder database via protocols such as Wi-Fi and Bluetooth, ensuring stable data transmission over long distances and in low-power environments. To guarantee real-time performance and transmission efficiency, binary stream files and the standardized JSON data format can be used to package the sensor data, and gzip compression can be applied before transmission to reduce bandwidth consumption.

[0031] Step 308: Based on preset processing rules, preprocess the acquired electronic medical record information data, the original data, and the sleep disorder data to obtain target data that meets the preset standards of the sleep disorder disease database. The preset processing rules include data cleaning rules and data noise reduction rules. The data cleaning rules are used to remove duplicate data and invalid data, and the data noise reduction rules are used to remove interference noise.

[0032] Step 310: Based on the established feature index, the target data is segmented and stored.

[0033] The feature index may include at least one of the following: a feature index established by sleep stages, a feature index established by sleep disorder types, and / or a feature index established by sleep events. The sleep stages and sleep disorder types are derived from the data analysis module or doctor annotations, and the sleep events may be derived from doctor annotations.

[0034] Feature indexes can be built according to sleep stages. The entire sleep process can be divided into five stages: Wake, N1, N2, N3, and REM sleep. Index functions can be defined. ,in, This indexing function can divide the original time series D(t) into five stages, each containing the target data within the corresponding time period.

[0035]

[0036] in, This indicates the start and end times of this sleep stage.

[0037] Feature indexes can be built according to sleep disorder types. Sleep disorder label types can include: insomnia, obstructive sleep apnea, narcolepsy, REM sleep behavior disorder, sleepwalking, restless legs syndrome, and periodic limb movement disorder. Index functions can be defined, assuming all physiological data of patient p are time series. Where t represents time, define the index function. ,in

[0038]

[0039] Mathematically, this can be represented as:

[0040] This indicates that the entire physiological data segment of patient p is classified as a specific sleep disorder type.

[0041] Feature indexes can be built based on sleep events, which may include micro-arousals, sleep apnea, limb movement, heart arrhythmias, blood oxygenation changes, environmental disturbances, etc. Therefore, the corresponding indexing function can be... ,in,

[0042] This indexing function helps to quickly find data within a period of time before and after a specific event. Mathematically, it can be represented as:

[0043] in, Indicates the point in time when the event occurred. and These represent the lengths of the time windows before and after the event, respectively.

[0044] It is understandable that feature indexes, such as those established by sleep stages, sleep disorder types, and / or sleep events, enable the rapid extraction of useful information from large amounts of data. For example, suppose a patient experiences multiple sleep apnea events at night. The doctor wants to understand the physiological responses before and after each apnea event, such as the trends in electrocardiogram (ECG) and electroencephalogram (EEG). The feature indexes described above can directly return ECG or EEG data for 30 seconds before and after the apnea event, thereby helping clinicians to more quickly assess the patient's sleep status during diagnosis and research. In one embodiment, the feature index may include at least one of the following: a feature index established by sleep stages, a feature index established by sleep disorder types, and / or a feature index established by sleep events.

[0045] In one embodiment, the sleep disorder database is equipped with an anomaly detection component, which is used to determine whether the sleep disorder data is abnormal, so as to remove the abnormal sleep disorder data.

[0046] In one embodiment, the sleep disorder specialty database is configured with an authentication component for authentication purposes to ensure the security of in-hospital data.

[0047] In one embodiment, the data storage module is divided into a first storage layer, a second storage layer, and a third storage layer based on the difference in access frequency. The first storage layer stores the raw data and sleep disorder data acquired in real time; the second storage layer stores the electronic medical record information data; and the third storage layer stores the target data. Specifically, the first storage layer stores the raw data and sleep disorder data acquired in real time, such as high-frequency physiological signals. The second storage layer stores the electronic medical record information data, such as multiple sleep records of the patient, treatment plans, and effects. The third storage layer stores the target data, such as processing results, diagnostic reports, and statistical data, for long-term archiving and data mining. To save costs, the above-mentioned storage division can use different storage particles. For example, the first storage layer can use SLC (Silicon Carbide) cache particles, the second storage layer can use MLC (Multi-Level Cell) particles, and the third storage layer can use TLC (Thin-Level Cell) particles. The third storage layer may even not use solid-state drive (SSD) storage particles as described above, and can use a hard disk drive (HDD) for storage. This effectively controls costs while ensuring storage efficiency.

[0048] In one embodiment, each storage node is further configured with a data standardization module. This module is used to standardize the target data based on a preset knowledge graph term embedding component, thereby classifying and storing the target data. The data standardization module can be used to standardize the target data based on a preset knowledge graph term embedding component, thereby classifying and storing the target data. Before data analysis, a large number of different relevant indicators are usually collected. Each indicator may have differences in its properties, dimensions, magnitude, usability, etc., making it impossible to directly use them to analyze the characteristics and patterns of the research object. When the levels of various indicators differ greatly, if the raw values ​​of the indicators are used directly for analysis, the role of indicators with higher values ​​in the comprehensive analysis will be amplified, while the role of indicators with lower values ​​will be weakened. Therefore, in order to better analyze the target data, the target data is classified and stored (e.g., insomnia, excessive sleepiness, sleep apnea, etc.) to facilitate the management of a sleep disorder-specific database.

[0049] Because various medical record texts may be written by different medical institutions and doctors, diagnostic information and terminology may often exhibit heterogeneity. In view of this, the data standardization module 212 described in this specification can be configured with a knowledge graph term embedding component to classify the target data, thereby achieving data standardization of textual information. Standardized text information categories can include the following: Symptoms, Diseases / Disorders, Lifestyle and Behavioral Factors, Physiological and Psychological Characteristics, Treatment Methods, Diagnostics and Monitoring, Medical Interventions and Surgery, Medications and Supplements, Risk Factors, Comorbid Symptoms and Complications, Biomarkers, Environmental Factors, Patient Demographics, Assessments and Scales, Genetic and Genomic Information, Data Types and Temporal Information, and Sleep Behaviors and Physiological Indicators.

[0050] Specifically, this knowledge graph word embedding component can be configured with a NER (Named Entity Recognition) model based on BERT (Bidirectional Encoder Representations from Transformers). This model can perform structured processing on medical record data, classifying each part of the text data as an entity into the specific categories mentioned above (i.e., the text information categories mentioned above). Then, it queries the correspondence of the entity in the constructed knowledge graph to map the entity to a standardized concept. For example, if the phrase "can't sleep well all night" appears in the medical record text, it will first be identified as the category "symptoms," and then matched to the node "insomnia" through our constructed knowledge graph.

[0051] The following instructions are combined with Figure 4 The specific text describes the above recognition process. Figure 4 This is a flowchart illustrating text processing based on a knowledge graph embedding component, provided in an exemplary embodiment of this specification.

[0052] Suppose the medical record text contains the following sentence: ; Step 402: The BERT-based NER model processes the text and identifies the symptom entity e: ; Here, E represents the extracted set of entities, where “cannot sleep at all during the first half of the night” is identified as a symptom entity e.

[0053] Step 404: After identifying entity e, convert entity e into a high-dimensional vector representation to facilitate matching with nodes in the knowledge graph.

[0054] After identifying entity e, entity e can be converted into a high-dimensional vector representation, which facilitates matching it with nodes in the knowledge graph.

[0055]

[0056] The high-dimensional vector representation is called a word embedding because, due to the characteristics of the BERT model, it can generate context-sensitive embeddings, making the meaning of each word more explicit in a given context. That is, the word embedding contains semantic information that includes the contextual information of the phrase "couldn't sleep at all during the first half of the night".

[0057] It should be noted that in knowledge graphs, node embeddings (such as "insomnia") are typically used to represent standardized medical terms. These terms are static embeddings without contextual information. Therefore, directly calculating the similarity between contextualized medical record description embeddings and context-free term embeddings in the knowledge graph is inappropriate. This specification therefore incorporates the BERT model to ensure consistency between term embeddings in the knowledge graph and contextualized embeddings in natural language descriptions. Specifically, this is achieved by generating term embeddings isomorphic to contextualized embeddings in natural language.

[0058] Step 406: Find the corresponding node in the knowledge graph through semantic matching.

[0059] To link the description "couldn't sleep at all during the first half of the night" in a medical record to a standard term in a knowledge graph, the corresponding node in the knowledge graph can then be found through semantic matching.

[0060] Suppose that the knowledge graph we construct contains embeddings of the following terms:

[0061] Specifically, the cosine similarity between the entity vector ve extracted from the medical record and the node vector in the knowledge graph can be calculated, and the node with the largest similarity is taken as the node with the highest similarity. Then, the node with the highest similarity is taken as the corresponding node in the knowledge graph as described above.

[0062]

[0063] Where vi∈V represents a node in the knowledge graph.

[0064] To ensure the correctness and consistency of the matching, after obtaining the similarity value, it can be further verified based on the edges in the knowledge graph. Specifically, this specification combines the similarity of the node word embedding vectors as described above with the semantic path in the knowledge graph to obtain an enhanced similarity (or final similarity) to obtain the final entity linking, accurately linking the description "completely unable to sleep in the first half of the night" in the entity medical record to the standardized node "insomnia".

[0065] The formula for the final similarity score described in this specification can be: ; Where α and β are weighting coefficients.

[0066] Based on the above steps, medical record text can be accurately matched to the corresponding nodes in the knowledge graph, thus achieving classification.

[0067] Furthermore, to further improve matching accuracy, this specification proposes a bidirectional matching strategy, which calculates not only the similarity between medical record descriptions and term embeddings but also the performance of terms in different contexts. For example, by retrieving multiple medical records related to "insomnia," the term embeddings are recalculated using their contexts. Specifically, multiple medical record descriptions related to term T ("insomnia") are retrieved, resulting in a context set C1, C2, ..., Cn, where each context Ci contains different descriptions of the term. For each context Ci, term T is placed within the context and encoded using a BERT model to obtain context-dependent term embeddings. : Right now, ; in, This indicates that the term T is embedded into the context Ci to form new input text, which is then encoded by the BERT model. [CLS] indicates that this is the semantic representation of the entire sentence.

[0068] For each context-dependent term embedding We can calculate its cosine similarity to the medical record description embedding ve: ; The final context-enhanced similarity score is obtained by averaging all context-related similarity scores. .

[0069] In one embodiment, any of the storage nodes is further configured with a data analysis module, which is used to analyze the target data based on a preset sleep disorder type analysis algorithm, a sleep stage analysis algorithm, and a similar sleep case recommendation algorithm, so as to classify the target data into sleep disorder types, classify the target data into sleep stages, and recommend similar sleep cases.

[0070] As mentioned earlier, common types of sleep disorders include insomnia, obstructive sleep apnea, narcolepsy, REM sleep behavior disorder, sleepwalking, restless legs syndrome, and periodic limb movement disorder. Accurate classification of sleep disorder types can better assist doctors in providing targeted treatment. The analysis algorithm in this module (Data Analysis Module 216) combines a rule-based judgment model based on the diagnostic and treatment guidelines for mental disorders with a multimodal deep learning model based on physiological-image-text, using a weighted average fusion to obtain the analysis results of sleep disorder types.

[0071] Specifically, firstly, the data analysis module 216 can be configured with a rule-based judgment model. This model uses parameters such as sleep latency, nocturnal awakenings, total sleep time, sleep efficiency, apnea index, apnea duration, changes in blood oxygen saturation, daytime sleepiness, abnormal behaviors during dreams, patient symptom reports, patient medical history, and family history to make decisions. For example, if a patient's number of nocturnal awakenings and apnea index exceed a certain threshold, and are accompanied by significant daytime sleepiness, the rule-based model will tend to diagnose obstructive sleep apnea.

[0072] Secondly, the data analysis module 216 can be configured with a multimodal deep learning model. This model can categorize data into three types: physiological data, imaging data, and text data, and then employ the most suitable deep learning technology for analysis in each category. For physiological data (such as electrocardiogram, electroencephalogram, electrooculogram, electromyography, blood oxygen saturation, and respiratory waveforms), a bidirectional long short-term memory network is used because it can capture the forward and backward temporal dependencies of physiological data, improving the understanding of complex time-series patterns. For imaging data (such as sleep video recordings, magnetic resonance imaging, and computed tomography scans), a channel attention residual convolutional neural network is used to emphasize important features and suppress irrelevant features, improving the feature extraction capability of imaging data. For text data (such as electronic health records, patient self-reports, and patient family medical histories), Transformer technology is used because it has a powerful global information capture capability and can better understand complex textual information. Subsequently, the deep learning model outputs a feature vector for each type of data. This feature vector serves as the input to a multilayer perceptron and is fused through a fully connected layer to learn the correlations and complementarities between different types of data.

[0073] Finally, a weighted average fusion strategy was used to combine the results of the rule-based judgment model and the multimodal deep learning model. Specifically, based on the performance of the two models on the training data, their respective weights were determined, and the outputs of both were combined using a weighted average method to obtain the sleep disorder type corresponding to the patient.

[0074] As mentioned earlier, the entire sleep process can be divided into five stages: Wake, N1, N2, N3, and REM sleep. Accurate sleep staging is crucial for treating sleep disorders. The data analysis module 216 can achieve sleep staging using a parallel dilated attention network model and medical rule correction. Specifically, the target data is first used as input to the model, with data divided into one sample every 30 seconds. Single-channel or multi-channel data can be selected. Next, a parallel convolutional neural network is used for feature extraction. Selecting A channels of data results in A parallel convolutional layers for feature extraction. For example, selecting single-channel EEG data and dual-channel EEG data would result in three parallel convolutional layers. Then, a concatenation layer is used to merge the results of different parallel convolutional layers along the feature channel dimension. Finally, the dilated convolution design and the QKV (Queries, Keys, Values) module in the attention mechanism are used for feature optimization. Dilated convolution can expand the receptive field without increasing computational cost, effectively capturing long-range dependency information. The attention-based QKV module can focus on important features, further improving the quality of feature representation. Finally, sleep staging is performed using a multilayer perceptron and the Softmax function. Specifically, after obtaining the sleep staging results, corrections and prompts are made based on medical rules to ensure the clinical accuracy and practicality of the results. For example, the transitions between certain sleep stages follow specific patterns; for instance, a direct transition from REM sleep to N3 sleep is uncommon and should be corrected. Furthermore, some sleep stages need to occur continuously for a certain period, such as N3, which typically lasts for at least several minutes. If the identified N3 stage is too short, correction may be necessary. The data analysis module 216 can automatically detect these anomalies and generate attention prompts, reminding doctors to carefully review this part of the data.

[0075] The recommendation of similar sleep cases can provide doctors with a reference for diagnosis and treatment. The recommendation algorithm configured in the data analysis module 216 learns the embedded representation of the data through a variational autoencoder, calculates similarity scores using Euclidean distance and cosine similarity, and uses a locality-sensitive hashing algorithm for efficient retrieval of similar cases. Specifically, for the structured data obtained by the data standardization module 212, the variational autoencoder compresses the data into a low-dimensional representation space, retaining key features and removing noise and redundant information, thereby obtaining a more representative low-dimensional data representation. After obtaining the low-dimensional representation, the algorithm calculates the similarity between different cases. For this purpose, the recommendation of similar sleep cases described in this specification uses two main similarity measurement methods: Euclidean distance and cosine similarity. Euclidean distance measures the similarity between two points by calculating the straight-line distance between them, and is suitable for measuring overall differences. Cosine similarity, on the other hand, reflects the directional similarity by comparing the angle between two vectors, and is suitable for capturing relative changes in data. The combination of these two methods can provide a more comprehensive and detailed similarity measurement. Next, the Locality Sensitive Hashing (LSH) algorithm maps similar cases to the same bucket using a set of hash functions, significantly reducing the number of cases that need to be compared and improving retrieval efficiency. Finally, the retrieved cases are ranked according to a comprehensive similarity score, and the most similar cases are recommended to doctors for reference.

[0076] Based on the above steps, a sleep disorder-specific database as described in this manual can be established, thereby maintaining relevant patient data.

[0077] The recommendation module 206 is used to input the patient's sleep assessment report, image analysis report and corresponding electronic medical record information into the sleep diagnosis and treatment model to obtain the diagnosis and treatment recommendation results corresponding to the diagnosis and treatment task.

[0078] In one embodiment, a patient's sleep assessment report can be generated based on the patient's sleep monitoring data; that is, the patient's sleep monitoring data is acquired to generate a corresponding sleep assessment report. Similarly, a corresponding image analysis report can be generated based on image data; that is, the patient's image data is acquired to generate a corresponding image analysis report. Specifically, sleep monitoring data may include electroencephalogram (EEG) data, electrooculogram (EOG) data, electromyogram (ECG) data, electromyogram (EMG) data, respiratory data, snoring data, blood oxygen saturation data, pulse rate data, body movement data, etc., while image data may include magnetic resonance imaging (MRI) data, computed tomography (CT) data, and positron emission tomography (PET) data, etc. The specific generation can rely on a preset model; that is, sleep monitoring data is input into a sleep assessment model to obtain a sleep assessment report, and image data is input into an image analysis model to obtain an image analysis report. The sleep assessment model may include a sleep staging model, a micro-arousal detection model, a body position detection model, and a snoring detection model, etc. The image analysis model may include a brain disease analysis model, an abnormal region annotation model, a structural brain injury detection model, and a functional brain activity analysis model, etc.

[0079] The sleep assessment report described in this instruction manual may include the patient's total sleep time, number of awakenings, sleep efficiency, latency time of N1 / N2 / N3 / REM sleep, duration of N1 / N2 / N3 / REM / Wake sleep, proportion of light sleep / deep sleep / REM sleep, microarousal index, apnea-hypopnea index, snoring frequency and intensity, abnormal blood oxygenation ratio, heart rate variability analysis results, periodic limb movement index, number of body movements, and other indicators and parameters.

[0080] The image analysis report described in this manual may include brain structural integrity assessment results, abnormal area indications, brain functional activity atlas descriptions, potential brain disease diagnostic references, descriptions of brain structural change trends, cerebral blood flow assessment results, and metabolic activity level assessment results.

[0081] The medical record data described in this instruction manual may include basic patient information, physical examination results, chief complaints, family medical history, lifestyle, psychological assessment results, routine laboratory test results, surgical treatment records, medication use records, sleep environment, and previous follow-up records.

[0082] To better match diagnostic and treatment tasks, this specification proposes a task-driven subgraph expansion strategy for generating auxiliary diagnostic content during the Retrieval Enhancement Generation (RAG) process. Specifically, based on a pre-defined sleep medicine knowledge graph, a temporary subgraph corresponding to the sleep assessment report, image analysis report, and corresponding medical record data can be determined. Based on the temporary subgraph and the pre-defined generation strategy, a contextual subgraph that conforms to the diagnostic and treatment task is generated. The contextual subgraph is then converted into a corresponding natural language description to obtain the target input data. In other words, after finding an approximate subgraph matching the patient's input information from the sleep medicine knowledge graph, a learnable model is constructed to define the node relationship requirements related to the diagnostic and treatment task according to different diagnostic and treatment tasks (such as diagnostic reference, treatment suggestions, lifestyle adjustment suggestions, etc.). This dynamically expands the nodes and edges related to patient information, constructing a task-specific contextual subgraph containing key information. Through this task-driven expansion, the generation model is ensured to generate accurate diagnoses or suggestions based on the most relevant context. It not only matches the patient's current symptoms but also incorporates related diseases, comorbidities, treatment plans, and the latest medical research findings, thereby improving the comprehensiveness and interpretability of the diagnostic content. The following section of this manual provides a detailed explanation of the knowledge graph matching strategy.

[0083] A sleep medicine knowledge graph can be represented as Let V represent the set of nodes and E represent the set of edges. The combination of the node set and edges is formed by fusing the following subgraphs: Disease Graph, Symptoms and Signs Graph, Diagnosis and Treatment Graph, Risk Factors Graph, Comorbidity Graph, Patient Management and Follow-up Graph, and Lifestyle and Behavior Graph. The fused node set V contains the nodes of all subgraphs. The edges in all subgraphs can be called original edges. During the fusion process, the edge set can be added from two new sources: 1. Edges added based on the relationships between nodes in different subgraphs, which can be called added edges. For example, new causal or correlational relationships may be established between disease nodes and symptom nodes in different subgraphs. 2. Edges generated by the inference engine that reflect the hidden relationships inferred between existing nodes and edges can be called inferred edges. The fused complete knowledge graph G' includes a node set V and a final edge set E': .in .

[0084] Patient input information Temporary subgraph constructed from the input This can include patient case data, sleep assessments, and other related information. Task set: Each task (Such as etiological analysis and treatment recommendations) This method constructs a learnable model to define its related node relationships. A subgraph constructed given patient input. It can be found in knowledge graphs Find the most similar subgraph , ,in It is similarity, calculated using cosine similarity based on the embedding vectors of nodes and edges.

[0085] Based on task type and user input subgraph Predict the set of relations to be used The formula is as follows:

[0086] in The model scoring function outputs a matching score for a given task, input subgraph, and set of relations. The parameters θ are learned during training to maximize this score.

[0087] Task-driven subgraph expansion strategy based on task The system expands the context nodes associated with the current subgraph node: The expanded set of nodes is as follows:

[0088] The expanded edge set consists of the edges of the original subgraph and the newly expanded edges:

[0089] Ultimately based on the task Generate a task-specific contextual subgraph from the approximate subgraph corresponding to the patient input: After template-based transformation, the context subgraph is converted into a natural language description that meets the context requirements. Subsequently, these natural language descriptions, along with the patient input information, are used as contextual input to the Retrieval Augmentation (RAG) model to generate task-relevant text information.

[0090] In one embodiment, to optimize the subgraph matching and expansion strategy, a matching strategy optimization based on Human Feedback Reinforcement Learning (RLHF) can be introduced. This involves acquiring feedback information from doctors and / or patients regarding the context subgraph, which is used to evaluate the accuracy of the generated context subgraph; and updating the generation strategy based on the feedback information to generate context subgraphs that conform to the diagnostic task. The subgraph matching and expansion strategy is continuously optimized through Human Feedback Reinforcement Learning (RLHF). During the operation of the RAG system, user or doctor interaction feedback can be recorded, and this feedback information can be used to update the model's generation strategy. Unlike traditional large language models (RLHF), the feedback described in this specification does not focus on user ratings of the generated text response, but rather on evaluating and optimizing the quality of the retrieved subgraphs and their expansions. This allows for the selection of subgraphs most relevant to the patient's condition and task requirements, thereby improving the accuracy of diagnostic suggestions. Specifically, the system, based on the task... Generated context subgraph Converted into natural language or simple recommendation summaries, these can be used to generate feedback for physicians' quantitative scoring. Used to evaluate the matching and expansion quality of subgraphs. Of course, implicit feedback can also be passively collected, such as inferring the accuracy of the generated content based on the actual treatment plan used by the doctor. Accumulated feedback can be used to update the matching strategy. The gradient ascent method is used to optimize the strategy parameters, making future subgraph matching more accurate. In related technologies, during the inference phase, the system loads all task-related parameters from the general module, even if the current task only requires a portion of them. This increases the time cost of model loading and wastes computational resources. Since the general module contains update weights for all tasks, these weights still occupy memory space even if some task updates are not used. This leads to inefficient resource allocation, especially in multi-task scenarios, significantly increasing device memory pressure. This results in redundant loading and memory waste. Therefore, the sleep diagnosis and treatment model described in this specification can be configured with multiple task plugins, allowing the loading of task plugins matching the diagnosis and treatment task. The target input data, the patient's sleep assessment report, image analysis report, and corresponding medical record data are input into the sleep diagnosis and treatment model loaded with the task plugins to obtain the corresponding diagnosis and treatment recommendation results for the diagnosis and treatment task. In other words, in different diagnosis and treatment scenarios, the system dynamically loads the task plugin corresponding to the current task based on the task type. During task execution, only parameter updates related to that task are loaded. Specifically, this can be implemented based on the LoRA module. This design employs a combination of instruction-tuning and low-rank adaptive tuning (LoRA), creating an independent LoRA module for each task. This modular structure enables the system to adapt efficiently to multi-task and dynamic knowledge update scenarios. Assume the weight matrix of the base model is... In carrying out the mission At that time, the system loads the corresponding LoRA module, in which and This is the low-rank adaptation matrix for this task. The updated model weights are: The LoRA update for each task only applies to a subset of layers of the base model, ensuring efficient optimization of model weights and reduced redundancy. Base weight matrix: Keep frozen, only adding task-specific LoRA updates during inference. Final effective weights. Only in the current task Used in reasoning.

[0091] The treatment recommendations described in this manual can be generated based on the user's treatment tasks. For example, assuming the treatment task is for diagnostic reference, the corresponding treatment recommendation could be: Based on the patient's sleep assessment report and imaging analysis report, the patient may have moderate sleep apnea syndrome (AHI 15 times / hour) and potential early signs of Alzheimer's disease (abnormal signal in the hippocampus, and decreased function in the right prefrontal cortex). Further neuropsychological testing and possible cerebrospinal fluid biomarker testing are needed to confirm the diagnosis of Alzheimer's disease. Assuming the treatment task is for treatment recommendations, the corresponding treatment recommendation could be: Treatment recommendations for sleep apnea syndrome include continuous positive airway pressure (CPAP) therapy. It is recommended that the patient use a CPAP device every night to improve sleep quality and oxygen saturation. For potential Alzheimer's disease, it is recommended to start cholinesterase inhibitors and have regular cognitive function assessments. Assuming the treatment task is for lifestyle modification recommendations, the corresponding treatment recommendation could be: It is recommended that the patient quit smoking and reduce alcohol consumption to improve overall health and sleep quality. Moderate aerobic exercise, such as 30 minutes of walking daily, is recommended to help control weight and improve cardiovascular health. Optimizing the sleep environment is also advised, reducing noise and light disturbances and ensuring the bedroom is quiet, dark, and cool. Assuming the treatment task is complication risk assessment, the corresponding recommendations could be: Sleep apnea syndrome, if left untreated, may increase the risk of cardiovascular diseases (such as hypertension and heart disease). Early signs of potential Alzheimer's disease should be monitored, as disease progression can lead to severe cognitive impairment and a decline in quality of life. Assuming the treatment task is treatment effectiveness assessment, the corresponding recommendations could be: Regular follow-up is recommended after starting CPAP therapy, including monthly sleep quality assessments and checks on CPAP device usage. For Alzheimer's disease, cognitive function assessments and imaging studies are recommended every 6 months to monitor disease progression and the effectiveness of medication. Assuming the treatment task is clinical pathway development, the corresponding recommendations could be: After initial diagnosis, patients are advised to return for a follow-up visit within one month to assess the initial effectiveness of CPAP therapy. Follow-up visits should be conducted at 3 and 6 months to assess overall treatment effectiveness and adjust the treatment plan. A comprehensive physical examination and sleep assessment are conducted annually to ensure early detection and treatment of any new health problems.

Claims

1. A digital diagnostic and treatment system integrating sleep center information, characterized in that, The system is equipped with a sleep disorder specialty database and a large-scale sleep diagnosis and treatment model. The sleep disorder database stores patients' electronic medical record information, raw data collected by medical devices, and sleep disorder data collected by terminal devices. The large-scale sleep diagnosis and treatment model recommends treatment outcomes based on the data stored in the sleep disorder database. The system includes: The determination module is used to determine the treatment task corresponding to the current sleep disorder treatment recommendation; The calling module is used to call the electronic medical record information data stored in the sleep disorder disease database corresponding to the current patient, the image data collected by medical devices, and the sleep monitoring data collected by terminal devices, and generate corresponding image analysis reports and sleep assessment reports. The recommendation module is used to input the patient's sleep assessment report, image analysis report and corresponding electronic medical record information into the sleep diagnosis and treatment model to obtain the diagnosis and treatment recommendation results corresponding to the diagnosis and treatment task.

2. The system according to claim 1, characterized in that, The diagnostic and treatment tasks include at least one of the following: Diagnostic references, treatment recommendations, lifestyle modification recommendations, complication risk assessments, treatment effectiveness evaluations, and / or clinical pathway development.

3. The system according to claim 1, characterized in that, The diagnostic and treatment tasks are selected by the patient or doctor through a task list.

4. The system according to claim 1, characterized in that, The recommendation module is specifically used for: Based on a pre-defined sleep medicine knowledge graph, a temporary subgraph is determined that corresponds to the sleep assessment report, image analysis report, and corresponding electronic medical record information data; Based on the temporary subgraph and the preset generation strategy, a context subgraph that conforms to the diagnosis and treatment task is generated; The context subgraph is converted into a corresponding natural language description to obtain the target input data; The target input data, the patient's sleep assessment report, image analysis report, and corresponding electronic medical record information are input into the sleep diagnosis and treatment model to obtain the diagnosis and treatment recommendation results corresponding to the diagnosis and treatment task.

5. The system according to claim 4, characterized in that, The system also includes: An update module is used to obtain feedback information from doctors and / or patients regarding the context subgraph, the feedback information being used to evaluate the accuracy of the generated context subgraph; Based on the feedback information, the generation strategy is updated to generate a context subgraph that conforms to the diagnostic and treatment task.

6. The system according to claim 1, characterized in that, The sleep diagnosis and treatment model is configured with multiple task plugins, and the recommendation module is specifically used for: Load the task plugin that matches the diagnostic and treatment task; The patient's sleep assessment report, image analysis report, and corresponding electronic medical record information are input into the sleep diagnosis and treatment model loaded with the task plugin to obtain the diagnosis and treatment recommendation results corresponding to the diagnosis and treatment task.

Images