Comprehensive rehabilitation service informatization integration method and system

Abstract

The invention relates to the technical field of medical informatization, in particular to a comprehensive rehabilitation service informatization integration method and system, and the method comprises the steps: obtaining multi-source heterogeneous patient rehabilitation data distributed in a hospital information system, a rehabilitation evaluation system, rehabilitation treatment equipment, wearable equipment and an electronic medical record; performing standardization processing on the multi-source heterogeneous patient rehabilitation data based on a preset unified rehabilitation data model; analyzing, cleaning and mapping the multi-source heterogeneous patient rehabilitation data into standardized patient rehabilitation data conforming to a unified rehabilitation data model in real time through a special data adapter deployed in a multi-protocol data interface layer; and by using a bidirectional synchronous data bus based on an event-driven architecture, issuing the standardized patient rehabilitation data to a specified rehabilitation device or client through the bidirectional synchronous data bus and an instruction generated by the central data warehouse. According to the invention, multi-source heterogeneous data from different systems and devices can be integrated to form a comprehensive, continuous and real-time health view for patients.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical informatization, in particular to a comprehensive rehabilitation service informatization integration method and system. BACKGROUND

[0002] Comprehensive rehabilitation service aims to provide comprehensive assessment, treatment and training for patients with functional disorders through multidisciplinary cooperation, so as to maximize the recovery of their physiological, psychological and social functions. With the development of information technology, the use of information technology to integrate rehabilitation resources, optimize service processes and improve rehabilitation results has become an important development direction in this field.

[0003] In the prior art, some information systems aimed at realizing personalized and dynamic adjustment of rehabilitation programs have appeared; however, the applicant has found in practice that there are still many technical problems to be solved in the prior art. First, rehabilitation data is widely distributed in hospital information systems, rehabilitation assessment systems, various rehabilitation treatment devices, wearable monitoring devices and electronic medical records. These systems have different standards and protocols, resulting in a serious "data island" phenomenon, which cannot form a comprehensive and continuous health view of the patient. Secondly, the "dynamic adjustment" of the existing program often relies on periodic follow-up or limited monitoring data, lacking comprehensive analysis capability of real-time, multi-dimensional data of patients in the rehabilitation process, resulting in lag in the adjustment of the rehabilitation program, making it difficult to realize true personalization and adaptive optimization. Finally, the evaluation of rehabilitation effect relies too much on subjective scales, which not only makes it difficult to guarantee the objectivity and consistency of the evaluation results, but also makes the evaluation dimension single and the sensitivity low, which cannot accurately capture the subtle functional changes of the patient and provide timely and effective quantitative feedback to the decision system. SUMMARY

[0004] The present application aims to provide a comprehensive rehabilitation service informatization integration method, comprising: acquiring multi-source heterogeneous patient rehabilitation data distributed in hospital information systems, rehabilitation assessment systems, rehabilitation treatment devices, wearable devices and electronic medical records; performing standardized processing on the multi-source heterogeneous patient rehabilitation data based on a preset unified rehabilitation data model, the unified rehabilitation data model defining data entities and standard formats of patient archives, functional assessment, physiological and motion data, rehabilitation intervention records and effect evaluation results covering the whole cycle of patient rehabilitation; through a special data adapter deployed in a multi-protocol data interface layer, the multi-source heterogeneous patient rehabilitation data is real-time parsed, cleaned and mapped into standardized patient rehabilitation data conforming to the unified rehabilitation data model; using a bidirectional synchronous data bus based on an event-driven architecture, the standardized patient rehabilitation data is real-time imported into a central data warehouse for storage and management, and the instructions generated by the central data warehouse are issued to the designated rehabilitation devices or clients through the bidirectional synchronous data bus, so as to realize the synchronization of data and instructions across systems.

[0005] By adopting the above scheme, the data island in the rehabilitation service is effectively broken, and multi-source heterogeneous data from different systems and devices can be integrated to form a comprehensive, continuous and real-time health view of the patient.

[0006] Optionally, the method further comprises dynamically generating and optimizing a personalized rehabilitation scheme using the standardized patient rehabilitation data, and the step comprises: constructing a rehabilitation knowledge graph comprising patient, disease, symptom, rehabilitation task and physiological indicator entities and their mutual relationships based on the standardized patient rehabilitation data; using a reinforcement learning model for dynamic decision-making, wherein a patient state vector generated based on the rehabilitation knowledge graph is used as the state of the reinforcement learning model, a task and parameters selected from a pre-defined rehabilitation task library are used as actions, and a reward signal generated based on a quantitative evaluation result of the rehabilitation effect is used as a reward; the reinforcement learning model selects an optimal action according to the current state to generate or adjust a personalized rehabilitation scheme, and continuously optimizes its decision-making strategy according to the reward signal obtained after executing the scheme.

[0007] By adopting the above scheme, the generation of the rehabilitation scheme no longer depends on static templates or lagging follow-ups, and the system can make intelligent decisions and continuously optimize itself according to the real-time state of the patient in multiple dimensions, recommend the most effective rehabilitation path for each patient at each stage of rehabilitation, and achieve precise rehabilitation for "thousands of people with thousands of faces".

[0008] Optionally, the step of constructing the rehabilitation knowledge graph comprises: defining patient, disease, symptom, injury site, rehabilitation task, rehabilitation device and physiological indicator as entities; and defining diagnostic relationships, functional dependency relationships, causal relationships between interventions and effects, and improvement or limitation relationships between the entities as edges connecting different entities, thereby converting unstructured rehabilitation data into a structured knowledge network.

[0009] By adopting the above scheme, the originally scattered and unstructured rehabilitation data is effectively organized into a structured knowledge network containing rich semantic information, providing high-quality input for the reinforcement learning model.

[0010] Optionally, the decision-making strategy of the reinforcement learning model is implemented through a deep neural network, which learns the mapping relationship from the patient state vector to the optimal rehabilitation task and parameters to maximize the long-term cumulative reward.

[0011] By adopting the above scheme, the powerful nonlinear fitting capability of the neural network enables it to learn and understand the extremely complex relationship between the patient state vector and the optimal rehabilitation task, thereby making more accurate and intelligent decisions than traditional rule engines or simple models, and the ultimate goal is to maximize the long-term cumulative rehabilitation benefits of the patient.

[0012] Optionally, the unified rehabilitation data model takes the International Classification of Functioning, Disability and Health framework as the top-level structure, and semantically associates physiological and motor data, functional assessment data, and rehabilitation intervention records.

[0013] By adopting the above scheme, the classification framework is an international standard for describing health and health-related conditions published by the World Health Organization, and adopting the framework can ensure that the model is aligned with international clinical standards at the semantic level, greatly improving the semantic association and interoperability of the data.

[0014] Optionally, it also includes multi-dimensional quantitative evaluation of the rehabilitation effect of the patient, and the steps include: synchronously collecting and fusing multi-source time series data of the patient during the execution of the rehabilitation task, the multi-source time series data at least including electromyographic signals, kinematic data and heart rate variability data; inputting the fused multi-source time series data into a pre-trained deep learning model; the deep learning model analyzes the multi-source time series data and outputs a comprehensive quantitative evaluation index for representing the current functional recovery level of the patient.

[0015] By adopting the above scheme, a multi-source time series data fusion analysis method based on deep learning is adopted to convert the traditional subjective scale evaluation into objective and precise quantitative index evaluation. This method not only accurately reflects the rehabilitation progress, but also provides interpretable analysis through the attention mechanism, revealing the key factors affecting the rehabilitation effect and providing a scientific basis for clinical intervention.

[0016] Optionally, the deep learning model is a model based on a long short-term memory network and an attention mechanism, wherein the long short-term memory network is used to learn the time dependence within each time series data stream, and the attention mechanism is used to identify key time points or key data indicators that have a significant impact on the evaluation result.

[0017] By adopting the above scheme, the deep learning model adopts a combination of long short-term memory networks and attention mechanisms; wherein the long short-term memory network is good at capturing and learning the time dependence within the time series data, such as the activation sequence of different muscles in a movement; and the attention mechanism can automatically identify and focus on the key time points or key data indicators that have the greatest impact on the evaluation result; for example, in an arm lifting training, the model can automatically focus on the joint angle data at a specific moment that causes the movement to fail.

[0018] Optionally, it also includes: using the weight distribution learned by the attention mechanism to identify and output physiological or kinematic factors that have a key impact on the comprehensive quantitative evaluation index, providing interpretable analysis for the rehabilitation effect.

[0019] By adopting the above approach and analyzing the weight distribution learned by the attention mechanism, the system can not only provide an evaluation score, but also identify and output the key physiological or kinematic factors that lead to the score. For example, the system can provide a specific analysis such as "the low score in this training session is mainly due to delayed deltoid muscle activation," which provides valuable clinical insights for rehabilitation therapists to adjust their programs.

[0020] Optionally, the comprehensive quantitative evaluation index or its change over time can be used as the reward signal for the reinforcement learning model, thereby constructing a closed-loop control system from data integration and effect evaluation to program optimization, and realizing continuous adaptive adjustment of personalized rehabilitation programs.

[0021] By adopting the above approach, comprehensive quantitative evaluation indicators (or their changes over time) are directly used as reward signals for the reinforcement learning model. This constructs a complete closed-loop control system from data integration and effect evaluation to program optimization. Each time a patient completes a training session, the system can objectively evaluate the effect and feed it back to the decision-making model as a positive or negative "reward," driving it to continuously optimize itself and thus achieve continuous adaptive adjustment of personalized rehabilitation programs.

[0022] The second objective of this application is to provide an integrated information system for comprehensive rehabilitation services, including a memory and a processor. The memory stores a computer program that can be loaded by the processor and executed by the aforementioned integrated information system for comprehensive rehabilitation services. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the integrated information technology method for comprehensive rehabilitation services in this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0025] like Figure 1 As shown in the figure, this application discloses a method for information integration of comprehensive rehabilitation services, which specifically includes the following steps.

[0026] S01: Acquire multi-source heterogeneous patient rehabilitation data distributed across hospital information systems, rehabilitation assessment systems, rehabilitation treatment equipment, wearable devices, and electronic medical records.

[0027] It is understandable that these inputs are raw data streams from different physical locations and logical systems, with varying formats and protocols. This embodiment can be implemented using a multi-protocol data interface layer. This interface layer pre-configures connectors and protocol parsers for different data sources, then establishes a stable data link and performs a protocol handshake, ultimately outputting raw data packets with source identifiers. Specifically, for hospital information systems and electronic medical records, the system establishes a connection with the hospital server through an adapter conforming to the Layer 7 protocol of the medical information exchange standard, using a Secure Sockets Layer encrypted channel, periodically polling or... The message middleware subscription method obtains basic patient information, such as the patient's unique identifier (a string), admission diagnosis (using the International Classification of Diseases, 10th Revision, e.g., "S72.0" represents a femoral neck fracture), allergy history (text description), and medical order records (including fields such as order identifier, order date, and content). For rehabilitation assessment systems, such as software used for Fogg-Meyer balance function assessment, the data may be stored in a proprietary binary format or comma-separated value files. The data interface layer of this application will deploy a file system monitor, which will detect new assessment results. Once the file is generated, its contents are immediately read and parsed to extract the assessment date, total score (integer, range 0-14), and scores for each sub-item. For rehabilitation therapy equipment, taking an upper limb rehabilitation robot as an example, this device broadcasts kinematic data packets to the local area network at a frequency of 60 times per second via its proprietary User Datagram Protocol. The data packet payload is a fixed-length byte array, where specific byte positions correspond to specific data. For example, bytes 4-7 represent the shoulder joint flexion-extension angle (32-bit floating-point number, unit: degrees, range: -30.0 to 180.0), and bytes 8-11 represent the elbow joint flexion-extension angle (32-bit floating-point number). Points, unit: degrees, range: 0.0 to 150.0), the user datagram protocol listener at the data interface layer captures these packets and attaches a device identifier and a high-precision timestamp; for wearable devices, such as a wireless electromyography (EMG) sensor, the device communicates with the data acquisition terminal via Bluetooth Low Energy (BLE) protocol, and its sampling frequency can be set to 1024 Hz. The Bluetooth Low Energy adapter at the interface layer will pair and connect with it, and subscribe to its EMG signal characteristic notifications, continuously receiving the EMG signal amplitude of each channel (16-bit signed integer, representing microvolts, range: -32768 to 32767).

[0028] S02: Based on the pre-set unified rehabilitation data model, the rehabilitation data of patients from multiple sources and heterogeneous structures are standardized. The unified rehabilitation data model defines the data entities and standard formats that cover the patient's records, functional assessments, physiological and motor data, rehabilitation intervention records and effect evaluation results throughout the entire rehabilitation cycle.

[0029] In this embodiment, the unified rehabilitation data model uses the International Classification of Functioning, Disability and Health (ICF-5) framework as its top-level structure, semantically associating physiological and motor data, functional assessment data, and rehabilitation intervention records. This ICF-5 is a standard published by the World Health Organization, providing a scientific framework for describing health and health-related conditions. By adopting this framework, the unified rehabilitation data model of this application achieves strong semantic interoperability. Specifically, the unified rehabilitation data model divides the data into several entities. For example, the "Patient Profile" entity contains demographic information (e.g., age, sex) and clinical diagnosis (represented by the ICF-51 code), and is associated with one or more question lists of this classification framework, such as "b730 Muscle Strength Function" impairment. The "Functional Assessment" entity stores various assessment results, and its structure includes the assessment scale name (e.g., "Fogg-Meyer Functional Assessment"), total score, assessment date, and a list of sub-items associated with the classification framework code. For example, the scoring item on upper limb motor ability in the assessment scale is mapped to the classification framework code "d445 Use of Hand and Arm". The "Physiological and Motor Data" entity is designed for storing high-frequency time-series data. Its structure includes session identifiers, timestamp sequences, and multiple data channels. Each channel explicitly defines the data type (e.g., electromyography signals, joint angles, heart rate), unit (e.g., millivolts, degrees, beats / minute), and source device identifier. Furthermore, each data channel is associated with one or more codes within this classification framework via metadata. For example, an electromyography signal channel collected from the deltoid muscle would be tagged and associated with "b730.2 shoulder muscle strength," while shoulder joint angle data from the upper limb rehabilitation robot would be associated with "b710 joint movement function." The "Rehabilitation Intervention Record" entity... This records the details of each treatment or training session, including the task name (e.g., "active shoulder flexion training"), equipment parameters (e.g., resistance set to 5 Nm), number of repetitions and sets. Similarly, each intervention task is mapped to one or more codes of this classification framework through its treatment objectives, such as "d445.1 lifting or carrying an object". Through this semantic association based on this international classification framework, previously isolated data points (such as a single electromyographic burst, a single joint movement, or a single rating scale score) are placed under a unified functional framework, achieving semantic integration across data types.

[0030] S03: Through a dedicated data adapter deployed in the multi-protocol data interface layer, multi-source heterogeneous patient rehabilitation data is parsed, cleaned, and mapped in real time into standardized patient rehabilitation data that conforms to a unified rehabilitation data model.

[0031] Specifically, during the parsing phase, the adapter converts binary or text data into key-value pairs that can be manipulated internally according to the data source's protocol specifications. For example, for the aforementioned upper limb rehabilitation robot user datagram protocol data packet, the adapter extracts the angle values ​​of joints such as the shoulder, elbow, and wrist based on predefined byte offsets and data types (e.g., 32-bit floating-point numbers). During the cleaning phase, the adapter performs a series of preprocessing operations to improve data quality. For example, for the acquired electromyographic signals, a fourth-order Butterworth bandpass filter with a passband range of 20 Hz to 450 Hz is applied to filter out motion artifacts and high-frequency noise. A 50 Hz notch filter is also applied to eliminate power frequency interference. For kinematic data, [further details are needed]. The adapter can use a Kalman filter to smooth the trajectory, reduce jitter, and estimate angular velocity and angular acceleration. If there are missing values ​​in the data, such as due to a brief interruption of the sensor signal, the adapter can use linear interpolation or previous value filling to complete the data, but it will also record a data quality mark to indicate that the data point has been interpolated. In the mapping stage, the adapter fills the cleaned data fields into the standardized structure of the unified rehabilitation data model. For example, it fills the parsed and cleaned shoulder joint angle time series data, along with its timestamp sequence, into an instance of the "physiological and motor data" entity, and associates it with the source device identifier, patient identifier, and the international classification framework code "b710", finally generating a structured data object. For example, an object instance that records kinematic time-series data contains multiple fields, such as: a "Patient Identifier" field with the value "P001"; a "Session Identifier" field with the value "Sess-abc"; a "Data Type" field with the value representing the kinematic time-series data; a "Functional Classification Code" field with the value being a list containing "b710"; a "Data Source" field, which contains a "Device Identifier" subfield with the value being the serial number of the rehabilitation robot; a "Timestamp Sequence" field, which is a list containing multiple timestamp values; and a "Value" field, which contains multiple subfields named after specific movements, such as "Shoulder Flexion-Extension Angle" and "Elbow Flexion-Extension Angle," each subfield having a value that is a list of values ​​corresponding to the length of the timestamp sequence.

[0032] S04: Utilizing a bidirectional synchronous data bus based on an event-driven architecture, standardized patient rehabilitation data is imported into a central data warehouse in real time for storage and management. Instructions generated by the central data warehouse are then sent to designated rehabilitation devices or clients via the bidirectional synchronous data bus to achieve cross-system data and instruction synchronization.

[0033] Specifically, this application embodiment can employ a distributed messaging platform as the data bus and define multiple topics to isolate different types of communication flows. For example, all standardized physiological and motor data generated by the data adapter will be published to a topic called "Rehabilitation Standardized Time Series Data," while functional assessment results will be published to the topic "Rehabilitation Standardized Assessment Data." One or more consumer services, such as a data warehouse write service, will subscribe to these topics. Once a new message arrives, it will be immediately retrieved and written to the corresponding database. For time series data, it is preferable to store it in a dedicated database optimized for processing time series data to optimize query performance. For structured patient files and assessment records, ... The data is stored in a relational database. The "other direction" of bidirectional synchronization is reflected in the issuance of instructions. When a subsequent intelligent decision-making module (such as the reinforcement learning model described below) generates a new rehabilitation task, it constructs an instruction event and publishes it to a specific instruction topic, such as the topic "Rehabilitation Instruction - Upper Limb Rehabilitation Robot". The payload of this instruction event is a structured data object that describes the task parameters in detail. Its fields include: "Instruction Identifier"; "Target Device"; "Task Name", such as the name "Shoulder Joint Resistance Abduction"; and the "Parameter" field, which contains multiple subfields, such as "Resistance Level", with a value of 5; "Number of Repetitions", with a value of 15; and "Number of Sets", with a value of 3. The data adapter deployed on the rehabilitation equipment or its control computer acts not only as a data producer but also as an instruction consumer. It subscribes to instruction topics corresponding to its equipment. Once it receives a new instruction message, it parses its content and converts it into specific operations that the equipment can execute through the equipment's software development kit or application programming interface, thereby controlling the equipment to start a new training task. This event-driven publish / subscribe model decouples the various components of the system, ensuring the real-time, reliable, and scalable transmission of data and instructions, and successfully constructing a dynamic, closed-loop information flow channel.

[0034] It is understood that the embodiments of this application may also include S05: dynamically generating and optimizing personalized rehabilitation plans using standardized patient rehabilitation data, specifically including the following steps.

[0035] S51: Based on standardized patient rehabilitation data, construct a rehabilitation knowledge graph that includes entities such as patients, diseases, symptoms, rehabilitation tasks, and physiological indicators, as well as their interrelationships.

[0036] Specifically, the steps to construct a rehabilitation knowledge graph include: defining patients, diseases, symptoms, injury sites, rehabilitation tasks, rehabilitation equipment, and physiological indicators as entities; defining diagnostic relationships, functional dependencies, causal relationships between interventions and effects, and improvement or limitation relationships between entities as edges connecting different entities, thereby transforming unstructured rehabilitation data into a structured knowledge network.

[0037] Understandably, the definition and extraction of entities (nodes) are as follows: Patient entity, with attributes including age, gender, height, and weight; Disease entity, uniquely identified by the International Classification of Diseases, 10th Revision (ICD-10), such as "I63.9" (cerebral infarction); Symptom / functional impairment entity, uniquely identified by the International Classification of Functioning, Disability and Health (ICF-10), such as "b760" (control of voluntary motor function); Rehabilitation task entity, such as "seated balance training - Level 3"; Rehabilitation equipment entity, such as the aforementioned upper limb rehabilitation robot; Physiological indicator entity, such as "integral electromyography value of the anterior deltoid muscle"; The definition and extraction of relations (edges) are as follows: For example, by parsing the diagnostic records in the electronic medical record, the diagnostic relation (Patient P001) - [diagnosed as] -> (Disease I63.9) can be extracted; by analyzing the mapping between the results of the functional assessment scale and the functional classification code, the relation (Disease I63.9) - [usually leads to] -> (functional impairment b760) can be established. Functional dependencies; through statistical analysis of a large number of historical rehabilitation intervention records and effect evaluation data, for example, through association rule mining, the causal relationship between intervention and effect of (rehabilitation task "seated balance training - level 3") - [improvement {confidence: 0.85, effect size: 0.3}] -> (functional impairment b760) can be found, where confidence and effect size are quantitative indicators; at the same time, improvement or limitation relationships can also be defined, for example (physiological indicator "integral electromyography value of anterior deltoid" > 0.5 mV*s) - [is a prerequisite for execution] -> (rehabilitation task "active shoulder flexion 90 degrees"), indicating that performing a certain task requires a certain muscle strength base. Through this process, the tabular data originally stored in the relational database is transformed into a semantically rich knowledge network, enabling the system to perform complex reasoning such as "starting from patient P001, finding their functional impairment, and then recommending rehabilitation tasks that may improve this impairment and meet their current physiological conditions".

[0038] S52: Dynamic decision-making is carried out using a reinforcement learning model, in which the patient state vector generated based on the rehabilitation knowledge graph is used as the state of the reinforcement learning model, the tasks and parameters selected from the predefined rehabilitation task library are used as actions, and the reward signal generated based on the quantitative evaluation results of rehabilitation effect is used as the reward.

[0039] The construction of the state requires a comprehensive and accurate description of the patient's current condition. In this embodiment, the patient state vector can be a dense vector with 384 dimensions. Its input is generated based on the aforementioned knowledge graph and real-time data, and its components include: 1) Demographic information (5 dimensions), such as age, gender, height, weight, and duration of illness, all of which are normalized; 2) Diagnostic information (100 dimensions), which converts the patient's main diagnosis (represented by the International Classification of Diseases, 10th Edition) into a 100-dimensional vector through a pre-trained embedding layer; 3) Functional state information (150 dimensions), which extracts the codes of all currently associated functional impairments of the patient and obtains their three most recent quantitative assessment scores (specifically calculated). (The calculation method will be explained in detail later) to form a time series, which is then encoded into a 150-dimensional vector through a small gated recurrent unit network to capture the trend of functional changes; 4) Knowledge graph embedding (129-dimensional): A graph embedding algorithm is used to learn the patient node in the knowledge graph to obtain a 128-dimensional vector. This vector encodes the patient's topological structure information in the entire knowledge network, which can represent its complex relationship with other patients, diseases, and tasks. Then, a unique code representing the current rehabilitation stage is concatenated, such as early / middle / late stage; The action space is defined as a discrete set, that is, from a predefined rehabilitation task library containing 128 different rehabilitation tasks. A task is selected. This rehabilitation task library is designed to cover activities for different body parts, difficulty levels, and training goals. For example, action number 5 might correspond to "upper limb robot-assisted shoulder joint horizontal adduction and abduction training, difficulty level 3, 15 repetitions / set." The output of the reinforcement learning model's policy network is the probability distribution of selecting these 128 actions. Reward is crucial for driving model learning, and its design directly relates to the final optimization goal. In this application, the reward signal originates from the quantitative evaluation results, and the final reward value is calculated based on a composite reward function. This function comprehensively considers multiple parts and performs a weighted summation using preset weights. The main part is the task itself. The changes in comprehensive quantitative evaluation indicators after execution are weighted as follows: a positive change brings a positive reward with a high weighting coefficient, such as 0.7; a secondary part is a penalty term, which represents the cost or risk of task execution. For example, if the heart rate variability index shows that the patient's physiological load is too high, or the electromyography data shows obvious muscle compensation patterns, this penalty term will take a negative value, such as -0.5, with a weight of 0.2; another part is the exploration reward, which is used to encourage the model to try new and less frequently chosen actions to avoid getting trapped in local optima, with a weight of 0.1. This composite reward function guides the model not only to pursue short-term effect improvement, but also to consider the patient's safety and comfort, and to maintain the exploratory nature of the strategy.

[0040] S53: The reinforcement learning model selects the optimal action based on the current state to generate or adjust a personalized rehabilitation plan, and continuously optimizes its decision-making strategy based on the reward signal obtained after executing the plan.

[0041] Specifically, the decision-making strategy of the reinforcement learning model is implemented through a deep neural network. This deep neural network learns the mapping relationship from the patient's state vector to the optimal rehabilitation task and parameters to maximize long-term cumulative rewards. This embodiment can use an advanced reinforcement learning algorithm—the proximal policy optimization algorithm. This algorithm is stable in training and has high sample efficiency. The model contains two core deep neural networks: a policy network (also called an execution network) and a value network (also called a review network). The structure of the policy network is as follows: the number of neurons in the input layer is 384, corresponding to the dimension of the state vector, followed by three fully connected hidden layers with 512, 256, and 256 neurons respectively, all using modified linear unit activation functions. A random deactivation layer (deactivation rate of 0.2) is added after each layer to prevent overfitting. The number of neurons in the output layer is 128, using a normalized exponential activation function, and the output is a vector representing the probability of selecting each rehabilitation task. The structure of the value network is similar to that of the policy network, but its output layer has only one neuron, using a linear activation function, to estimate the current value. The value of a previous state is the expected value of the cumulative reward that can be obtained by following the current policy starting from the current state. During training, the agent samples the probability distribution output by the policy network to select actions. After execution, it obtains a reward and the next state from the environment. The proximal policy optimization algorithm evaluates whether the selected action is good or bad relative to the average performance of the current state by calculating the advantage function. Then, it uses a pruned surrogate objective function to update the weights of the policy network. This objective function limits the step size of each update by pruning the probability ratio, thereby avoiding drastic policy changes and ensuring the stability of training. The value network is updated by minimizing the mean squared error between its predicted value and the actual observed reward. Through iterative training on a large number of real or simulated "state-action-reward-new state" sequences, this deep neural network-based proximal policy optimization model can gradually learn the mapping policy from complex patient states to the optimal rehabilitation task. Its goal is to maximize long-term cumulative rewards, thereby achieving continuous, intelligent, and adaptive optimization of the rehabilitation program.

[0042] The embodiments of this application may also include S06: performing multi-dimensional quantitative evaluation of the patient's rehabilitation effect, specifically including the following steps.

[0043] S61: Simultaneously collect and fuse multi-source time-series data of patients during the performance of rehabilitation tasks. The multi-source time-series data includes at least electromyographic signals, kinematic data, and heart rate variability data.

[0044] Specifically, assuming the patient is performing an active arm-raising exercise, the system can simultaneously activate multiple sensors for data acquisition: 1) Electromyography (EMG) signals: Using an 8-channel wireless EMG system, sensors are attached to the anterior, middle, and posterior deltoids, pectoralis major, biceps brachii, triceps brachii, latissimus dorsi, and upper trapezius muscles, with a sampling frequency of 1080 Hz, to acquire raw EMG signals; 2) Kinematic data: Using an inertial measurement unit (IMU) system, sensors are fixed to the upper arm and forearm, with a sampling frequency of 100 Hz, and the three-dimensional kinematic angles (Eulerian angles or quaternions) of the shoulder and elbow joints are calculated in real time using its built-in Kalman filter algorithm; 3) Heart rate variability data: Using a heart rate strap, the time interval between the peak R-waves representing ventricular depolarization during continuous heartbeats is recorded, with a sampling frequency of approximately 1-2 Hz, to assess the activity and physiological load of the autonomic nervous system; The first step in data fusion is temporal synchronization. First, the system uses the master clock of the device with the highest sampling rate (e.g., the electromyography system) as a reference, and performs precise time synchronization of all acquisition devices through a network time protocol or hardware trigger signal, with the error controlled within 1 millisecond. Next, the acquired data is preprocessed and aligned. The electromyography signal is bandpass filtered and root mean square calculated (window size 100 milliseconds, overlap 50 milliseconds) to obtain a smooth electromyography activation envelope, and downsampled to 100 Hz. Kinematic data and heart rate variability data are upsampled to 100 Hz through linear interpolation, so that all data streams have the same time axis and sampling rate. Finally, these aligned data streams are concatenated along the feature dimension to form a multi-channel time-series data matrix. For example, for a training movement lasting 10 seconds, at a sampling rate of 100 Hz, a matrix with shape (1000, 15) will be generated. The matrix, where 1000 is the time step and 15 is the number of features, namely 8 electromyographic signal channels + 6 kinematic degrees of freedom + 1 heart rate variability index, is the standard format for input into subsequent deep learning models.

[0045] S62: Input the fused multi-source time series data into a pre-trained deep learning model.

[0046] S63: The deep learning model analyzes multi-source time-series data and outputs a comprehensive quantitative assessment index to characterize the patient's current functional recovery level.

[0047] Understandably, deep learning models are based on Long Short-Term Memory (LSTM) networks and attention mechanisms. The LTM network learns the temporal dependencies within each time-series data stream, while the attention mechanism identifies key time points or key data indicators that significantly impact the evaluation results. The specific architecture of the deep learning model is as follows: The input layer receives a (1000, 15) data matrix, which is first standardized by a batch normalization layer. Then, the data is fed into a core network consisting of two stacked bidirectional LTM layers, each containing 128 hidden units. Using a bidirectional network means the model can simultaneously read sequence information from both forward and reverse directions, thus capturing temporal dependencies more comprehensively. For example, in a single arm-raising motion, the deep learning model can not only learn that "the deltoid muscle activates before the biceps brachii," but also associate it with a future-information-dependent pattern such as "the antagonist muscle failed to relax in time" at the end of the movement. The output of the two layers is a (1000, 256) shape. The sequence is given, where 256 is the dimension of the concatenated forward and backward hidden units. Next, this output sequence is fed into an attention mechanism layer, which calculates an attention weight for each of the 1000 time steps. The calculation typically involves: first, passing the hidden states of the Long Short-Term Memory network through a small feedforward network to obtain an alignment score; then, applying a normalized exponential function to the alignment scores of all time steps to obtain normalized weights. These weights represent the degree of attention the model pays to each time point in the action when making its final decision. Subsequently, a single, fixed-length context vector is obtained by weighted summing the attention weights with the hidden states of the Long Short-Term Memory network. The 256-dimensional context vector condenses the most critical information in the entire action sequence. Finally, this 256-dimensional context vector is fed into an output layer consisting of two fully connected layers with 64 and 1 neurons respectively. The final output is compressed to between 0 and 1 using a sigmoid activation function and then multiplied by 100 to obtain the final comprehensive quantitative evaluation index score. This deep learning model is trained through supervised learning on a large dataset of "time-series data - expert ratings" pairs, where expert ratings serve as the standard. For example, the average score of 0-100 given independently by three senior rehabilitation therapists to video recordings. The deep learning model learns network parameters by minimizing the mean squared error between its predicted score and the expert rating.

[0048] It is understood that, in order to enhance the clinical application value of the assessment results, the embodiments of this application may further include: using the weight distribution learned by the attention mechanism to identify and output physiological or kinematic factors that have a key impact on the comprehensive quantitative assessment indicators, so as to provide interpretability analysis for rehabilitation effects.

[0049] Specifically, when the deep learning model outputs a comprehensive quantitative evaluation score (e.g., 75.2 points), the system can simultaneously extract the corresponding attention weight vector. First, peak detection is performed on this weight vector to identify the N time points with the highest weight values, for example, N=3. These time points represent the key moments that the model considers most decisive for the final score. Then, the system backtracks to the input data matrix and extracts the values ​​of each feature channel at these key time points, such as electromyographic signals and joint angles. Next, these values ​​are compared with a preset database of "normal mode" or "ideal mode." For example, suppose that at a certain weight peak moment, the deep learning model finds that the electromyographic activation value of the anterior deltoid (feature channel 1) is far below its expected level, and... The system detected an unusually high activation value in the upper trapezius (feature channel 8), leading to the inference of a malfunctioning movement pattern called "scapular girdle compensation." Finally, the system integrated these analytical results into a human-readable text report, such as: "Overall score for this 'active arm raise' training: 75.2 points. Analysis shows that the low score is mainly due to a significant scapular girdle elevation compensation pattern in the mid-movement phase (approximately 3.4 seconds), characterized by insufficient activation of the anterior deltoid and excessive involvement of the upper trapezius. It is recommended that subsequent training focus on strengthening scapular stability control and guiding the patient to concentrate more on using the shoulder muscles." This interpretable analysis based on attention mechanisms transforms an abstract assessment score into concrete and actionable clinical guidance, greatly enhancing the system's practical value.

[0050] In this embodiment of the application, a comprehensive quantitative evaluation index or its change over time can also be used as a reward signal for the reinforcement learning model, thereby constructing a closed-loop control system from data integration and effect evaluation to program optimization, and realizing continuous adaptive adjustment of personalized rehabilitation programs.

[0051] Specifically, at time step t, the proximal strategy optimization model selects an optimal action based on the patient's current state vector through its policy network. For example, it decides to perform "upper limb robot-assisted shoulder joint horizontal adduction and abduction training, difficulty level 3". This action is constructed as an instruction and sent to the designated rehabilitation robot through a bidirectional synchronous data bus.

[0052] The patient performs tasks on the rehabilitation robot. During the process, multi-source time-series data of the patient are collected simultaneously, including electromyographic signals, kinematic data, heart rate variability data, etc., and these data are standardized. After the task is completed, a deep learning model based on long short-term memory network and attention mechanism analyzes the data and outputs a comprehensive quantitative evaluation index of the performance of this task, such as 82.5 points.

[0053] After obtaining the evaluation index, the system calculates its change relative to the previous evaluation result (e.g., 81.0 points), resulting in a change of +1.5. Simultaneously, the system calculates a cost penalty term based on the compensation pattern detection results in the heart rate variability data and kinematic data. If the patient's heart rate is stable and compensation is not obvious in this task, then this penalty term is 0. According to the composite reward function, the index change, cost penalty term, and exploration reward term (assuming this action is not frequently selected and a small exploration reward is obtained) are weighted and summed to calculate the final reward value obtained for this decision.

[0054] The reinforcement learning agent receives the reward value and the updated patient state after task execution, forming a complete experience unit containing the current state, the action performed, the reward obtained, and the next state to be entered, and stores it in the experience replay buffer. The proximal policy optimization algorithm samples a batch of experience data from the buffer, calculates the advantage function, and updates the parameters of its policy network and value network. Through this update, the policy network will be more likely to choose actions that bring positive changes in indicators and have no physiological risks when encountering similar states in the future.

[0055] The system enters the next time step t+1. The proximal policy optimization model makes a new decision based on the updated policy and the new state. For example, since the previous task was performed well (the reward was positive), the model may choose a more difficult task, such as "upper limb robot resistance shoulder joint horizontal adduction and abduction training, difficulty level 4". Then, the entire "decision-execution-evaluation-reward-learning" cycle starts again.

[0056] Through this closed-loop mechanism, the system achieves true adaptive optimization. It no longer relies on manually set rules or templates, but learns from objective quantitative assessment results by continuously interacting with the patient's real performance. It autonomously discovers the most effective rehabilitation path for a specific patient at a specific stage. If a rehabilitation task continuously leads to a decline or stagnation in the assessment score, the reinforcement learning model will gradually reduce the probability of selecting that task through the accumulation of negative rewards, and instead explore other potentially more effective tasks. Conversely, tasks that can continuously improve the patient's functional level will be reinforced.

[0057] This application also discloses a comprehensive rehabilitation service information integration system, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed by the aforementioned comprehensive rehabilitation service information integration method.

[0058] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0059] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0060] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0061] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0062] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0063] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. An information-based integrated method for comprehensive rehabilitation services, characterized in that, The method comprises: acquiring multi-source heterogeneous patient rehabilitation data distributed in a hospital information system, a rehabilitation assessment system, a rehabilitation treatment device, a wearable device, and an electronic medical record; standardizing the multi-source heterogeneous patient rehabilitation data based on a preset unified rehabilitation data model, the unified rehabilitation data model defining data entities and standard formats covering patient archives, functional assessment, physiological and motion data, rehabilitation intervention records, and effect assessment results in the whole cycle of patient rehabilitation; real-time parsing, cleaning, and mapping the multi-source heterogeneous patient rehabilitation data into standardized patient rehabilitation data conforming to the unified rehabilitation data model through a special data adapter deployed in a multi-protocol data interface layer; storing and managing the standardized patient rehabilitation data in a central data warehouse through a bidirectional synchronous data bus based on an event-driven architecture, and issuing instructions generated by the central data warehouse to designated rehabilitation devices or clients through the bidirectional synchronous data bus to realize synchronization of data and instructions across systems.

2. The method of claim 1, wherein, The method further comprises dynamically generating and optimizing a personalized rehabilitation plan using the standardized patient rehabilitation data, the steps comprising: constructing a rehabilitation knowledge graph containing patient, disease, symptom, rehabilitation task, and physiological indicator entities and their mutual relationships based on the standardized patient rehabilitation data; adopting a reinforcement learning model for dynamic decision-making, wherein a patient state vector generated based on the rehabilitation knowledge graph is used as the state of the reinforcement learning model, a task and parameters selected from a pre-defined rehabilitation task library are used as actions, and a reward signal generated based on a quantitative assessment result of rehabilitation effects is used as a reward; the reinforcement learning model selects an optimal action based on the current state to generate or adjust a personalized rehabilitation plan, and continuously optimizes its decision-making strategy based on the reward signal obtained after executing the plan.

3. The method of claim 2, wherein, The step of constructing the rehabilitation knowledge graph comprises defining patients, diseases, symptoms, injury sites, rehabilitation tasks, rehabilitation devices, and physiological indicators as entities, and defining diagnostic relationships, functional dependency relationships, causal relationships between interventions and effects, and improvement or limitation relationships between the entities as edges connecting different entities, thereby converting unstructured rehabilitation data into a structured knowledge network.

4. The method of claim 2, wherein, The decision-making strategy of the reinforcement learning model is implemented through a deep neural network that learns the mapping relationship from the patient state vector to the optimal rehabilitation task and parameters to maximize the long-term cumulative reward.

5. The method of claim 1, wherein, The unified rehabilitation data model takes the International Classification of Functioning, Disability and Health framework as the top-level structure, and semantically associates physiological and motion data, functional assessment data, and rehabilitation intervention records.

6. The method of claim 2, wherein, The method further comprises multi-dimensionally quantitatively assessing patient rehabilitation effects, the steps comprising: synchronously collecting and fusing multi-source time series data of patients during the execution of rehabilitation tasks, the multi-source time series data at least including electromyographic signals, kinematic data, and heart rate variability data; inputting the fused multi-source time series data into a pre-trained deep learning model; The deep learning model analyzes the multi-source time series data and outputs a comprehensive quantitative evaluation index representing the current functional recovery level of the patient.

7. The method of claim 6, wherein, The deep learning model is a model based on a long short-term memory network and an attention mechanism, wherein the long short-term memory network is used to learn the time dependence within each time series data stream, and the attention mechanism is used to identify key time points or key data indicators that have a significant impact on the evaluation result.

8. The method of claim 7, wherein, Further comprising: Using the weight distribution learned by the attention mechanism, identifying and outputting physiological or kinematic factors that have a key impact on the comprehensive quantitative evaluation index, thereby providing an interpretable analysis of the rehabilitation effect.

9. The method of any one of claim 6, wherein, Using the comprehensive quantitative evaluation index or its change over time as a reward signal for the reinforcement learning model, thereby constructing a closed-loop control system from data integration and effect evaluation to scheme optimization, and realizing continuous adaptive adjustment of the individualized rehabilitation scheme.

10. An information-based integrated system for comprehensive rehabilitation services, characterized in that, A memory and a processor, the memory storing a computer program capable of being loaded and executed by the processor to perform the comprehensive rehabilitation service informationization integration method according to any one of claims 1 to 9.

Images