A static-dynamic feature fusion framework based on long short-term memory and collaborative filtering

Abstract

Focusing on the recommendation and decision support module in management systems, this framework takes users’ basic information (e.g., demographics, cause-related attributes) and auxiliary data (e.g., temporal activity records, functional assessment data) as inputs. A loss function is employed to optimize model parameters, enabling the deep learning model to capture temporal patterns in service processes and mine key factors for training program decision-making. Methodologically, the framework integrates weighted matrix factorization to extract latent features from users’ static attributes while leveraging Long Short-Term Memory (LSTM) networks to model temporal dependencies from sequential auxiliary data, achieving a static-dynamic feature fusion that enables more precise personalization. By fusing LSTM's capability to process sequential data with collaborative filtering's strength in mining user-item correlations, the model not only achieves accurate personalized training program recommendations but also provides data-driven decision support for staff to formulate adaptive plans. Experimental validation demonstrates that the proposed framework enhances the precision and timeliness of training program recommendations, strengthens decision-making efficiency for staff, and improves the effectiveness of target service outcomes. Moreover, the model offers quantifiable predictions of functional outcomes alongside recommended programs, providing staff with actionable references for decision-making.

Introduction

In the context of intelligent management system development, the formulation of training programs in complex service scenarios (e.g., skill training, functional improvement) often relies on staff's subjective experience, leading to problems such as low decision-making efficiency, poor personalization, and difficulty in adapting to dynamic changes in user status. To address these challenges, intelligent recommendation combined with data-driven decision support has become a core demand for the upgrading of management systems—by integrating sequential data analysis and user-demand mining, it can help staff quickly identify optimal training schemes and improve the scientificity of decision-making. Among existing technical paths, Long Short-Term Memory (LSTM) excels at capturing temporal trends in time-series data (e.g., changes in user performance over time), while collaborative filtering effectively mines potential correlations between users and training items; their fusion provides a feasible solution for building high-performance recommendation and decision support models. Especially, the rehabilitation process of patients is a typical long-term dependence on time series: the breakthrough of a key motor function indicator (such as Ueda score) is often based on the results of weeks or even months of continuous training. The LSTM network, with its unique “forget gate” and “input gate” structures, can effectively learn and remember this long interval key information, avoiding the gradient vanishing problem of traditional Recurrent Neural Networks (RNNs), making it more suitable than standard RNNs for capturing the dynamic evolution of patient states. In contrast, although more advanced architectures such as Transformer perform equally well in long sequence modeling, they have higher computational overhead and are prone to overfitting on small-scale medical time-series datasets. Therefore, LSTM has become the preferred choice for this study due to its excellent performance and efficiency balance.

Against this backdrop, with the rapid economic development and deepening of healthcare reform, the rehabilitation medical sector—an important application field of management systems—has received growing attention.^1,2 Due to the continuous increase of patients with chronic diseases and stroke, the rehabilitation stage after treatment also poses challenges for medical institutions, such as higher public standards for rehabilitation services and the need for more scientific training program decision-making.^3–6 At present, the number of people with physical disabilities accounts for 29% of the total; physical disabilities are mainly caused by accidents, diseases, stroke, and other conditions that lead to nerve damage. The first step after limb injury is to reconstruct limb function through surgical treatment, while the body's nervous feedback function requires gradual recovery with the help of professional and scientific rehabilitation training methods—making the accuracy and adaptability of training program decisions critical to recovery outcomes.

At the same time, with the continuous innovation and development of computer network technology, internet technology is gradually participating in the informatization construction of China's medical system.^7–10 Patient information in hospitals and medical data generated during treatment require continuous informatization management; thus, medical information platforms based on various Internet technologies are emerging in large numbers. So far, information management in China's medical system has been increasingly applied in actual diagnosis and treatment processes, greatly improving the management level and work efficiency of medical institutions. The informatization management of medical rehabilitation systems is gradually being applied to major rehabilitation medical institutions and rehabilitation departments of regular hospitals,^11–13 aiming to manage rehabilitation institutions more efficiently and accelerate the pace of patient treatment to the greatest extent.

Traditional rehabilitation treatment could only be carried out through simple rehabilitation training and repeated follow-up examinations, leading to slow formulation of training plans and inability to store and effectively analyze treatment data. In common clinical practice, each step required constant adjustment and completion by staff from various departments—relying heavily on manual experience and resulting in inconsistent decision quality. With the popularization of informatization, different user permissions can be set in rehabilitation information systems to realize information management for managers in different departments. Therefore, combined with the management needs of hospital rehabilitation centers, research on rehabilitation center information management systems can not only meet the basic management needs of rehabilitation centers but also establish a more comprehensive digital management system. By using data analysis to support training program selection and decision-making, it can greatly reduce the work intensity of treating doctors, improve the accuracy of treatment decisions, and ultimately promote the intelligent upgrading of rehabilitation management.

The main innovation of this study lies in: firstly, at the methodological level, a static dynamic feature deep fusion framework is proposed. This framework utilizes weighted matrix factorization to extract latent features from patient static information, such as etiology and age, and combines them with state features extracted by LSTM networks from dynamic temporal data, such as functional evaluation history and training methods, to achieve accurate characterization of patient personalized needs. Secondly, at the application level, this type of deep collaborative filtering model has been systematically applied for the first time in the training program recommendation and decision support of rehabilitation medical management systems. The model not only outputs personalized training programs, but also predicts the functional evaluation values after program execution, providing quantifiable prognostic references for clinical physicians and significantly improving the objectivity and efficiency of decision-making.

Related technologies

The core content of this study is the training program recommendation system in the rehabilitation training process. The recommended program is obtained through the relevant recommendation algorithm, and the patient is provided with accurate training guidance, which helps them to train more efficiently. This section mainly introduces the related theoretical concepts of traditional recommendation algorithms and recommendation technology. It then combines with a deep learning network to extend to the Collaborative Filtering^14–16 based on LSTM Recurrent Neural Network^17–20 adopted by this system. Finally, the data simulation experiment is carried out by building a data model. Through the test set, the technical analysis results of the recommendation algorithm are used to verify the effect of the algorithm.

Traditional recommendation

Traditional recommendation algorithms can predict and analyze the potential needs of users to obtain recommendation results, thereby recommending similar items to users. Traditional recommendation algorithms are divided into two categories one is Content-based Recommendations (CB) ²¹ and the other is Collaborative Filtering Recommendations (CF), ²² whose principle is to use similar users or items for assistance recommendations.

Content-based recommendation

The first proposed algorithm is Content-based Recommendations, ²³ which recommends similar products to users based on their once-favorite products (collectively referred to as items in this article). An information recommendation system can recommend information based on the information frequently retrieved by the user. The recommendation process of content-based recommendations is divided into three steps: (1) Item representation: select the features of an item that can represent a user's past likes and dislikes (the content of the item); (2) Feature learning: learn the likes and dislikes of the user according to the feature data; and (3) Content recommendation: The features obtained in step 2 are used to match a group of items with the highest similarity for the user to complete the recommendation.

The basis of The Content-based Recommendations is to calculate the distance between a User (“U”) and the matching Item (“I”) through cosine similarity. The larger the value, the higher the probability that U likes I. Assuming that each user U and each item I can be represented as a k-dimensional feature vector, where a denotes the a-th feature dimension, the similarity is calculated as follows:

$$\cos (\mathrm{U},\mathrm{I})={\displaystyle \frac{\sum {\mathrm{U}}_{\mathrm{a}}^{*}{\mathrm{I}}_{\mathrm{a}}}{\sqrt{\sum {\mathrm{u}}_{\mathrm{a}}^2*}\sqrt{\sum {\mathrm{I}}_{\mathrm{a}}^2}}}$$

(1)

The advantages of Content-based Recommendations are as follows: first, it is very independent among users, and it is recommended only according to the user's likes and dislikes without affecting others; second, it is convenient to display the recommended results to users; third, the product library can be continuously updated Expansion, the probability of being recommended is the same. At the same time, it also has the following shortcomings: first, the core of the algorithm is to extract the abstract features of the items that the user likes, which leads to low accuracy; then, since user preferences may change, continuing to recommend content consistent with their original interests will inevitably lead to limitations; finally, due to the defect of Cold Start, it cannot meet the needs of new users.

Collaborative filtering

Model-based Collaborative Filtering is commonly used in practical applications, which use Matrix Factorization model to learn similar information between user and item. The latent Factor Model (LFM) and Matrix Factorization Model have been used most frequently in recent years. And they both complete the matrix through the dimensionality reduction method, which is also their same essential feature. The offline time complexity of LFM is O(K*S*F), where F means that there are F implicit representation classes, and S means that the model needs to be iterated S times. When LFM recommends, it needs to calculate the weight of the user's interest for all items and returns the item with the most significant weight after comparison. The disadvantage of this algorithm is that the recommendation list generation speed is too slow to perform real-time recommendations, and it can only be used for offline calculation of stored data. The Matrix Factorization Model needs to factorize the user-item matrix and recommend products to users by imputing missing values of the matrix. The Classic Collaborative Filtering has apparent limitations, and the Matrix Factorization Model can strengthen the factorization ability of a sparse matrix.

Compared with Content-Based Recommendations, Collaborative Recommendations are relatively simple, which reduces the extraction of preference features for complex users. It only needs to recommend based on user behavior data and does not require a complete labeling analysis of users or items. It also avoids label construction for some concepts that are difficult to quantitatively describe and discovers the potential interests and preferences of users well.

Deep learning networks

Compared with traditional recommendation algorithms, deep learning networks have many advantages when applied to recommendation systems, such as automatic extraction of data features, prediction of form diversity, improved interface in high-level time series, and better ability in feature extraction of text or images. Therefore, the probability of application of deep learning networks in information recommendation is significantly increased. Undoubtedly, the more deep learning is applied, the more complex the data will be. In practice, there will still be problems with reproducibility and benchmarks that need to be continuously improved. In this section, we will introduce the logical explanation of the neural network algorithm models commonly used in deep learning algorithms.

Recurrent neural network

As the name suggests, RNN is a kind of neural network with cyclic data.^24–33 The network model can process those sequential data, which means that the network output at a certain moment is not only related to the input at the current moment but also associated with the output at the previous moment. In recent years, RNN has been used in many sequential problems and is gradually used in recommendation systems. In recent years, RNN has been used in many sequential problems and is gradually used in recommendation systems.

Compared with Full Connect Neural Network, RNN adds a memory unit to save the state of the hidden layer. At a certain moment, the memory unit's content will enter the cyclic input process together with the input.

Long short-term memory recurrent neural network

The standard model of RNN is only suitable for short-term memory networks. To solve the Long-Term Dependencies problem of the network, the LSTM network appeared, which improved the RNN hidden layer algorithm.^25–28 The standard RNN hidden layer has only one h state, utterly dependent on the short-term input. The LSTM hidden layer adds a cell state (abbreviated as c) to save the long-term input state and store long-term memory.

Rehabilitation training program recommendation

After understanding the basic principles, the Rehabilitation training program recommendation module in this study uses Collaborative Filtering based on the LSTM Recurrent Neural Network as the recommendation model. This section will focus on the data processing in the model-building process, the specific steps of the training algorithm, and the analysis of the prediction results obtained under the model.

Deep learning platform

TensorFlow is an open-source platform for machine learning. The platform provides a stable API for Python data analysis. The capabilities and most essential features of Tensorflow are summarized in the following three points.

High-performance numerical calculation library: It can be used by various application programs only by importing functions. It is written in C++ and provides interfaces for different programming languages. One of the most complete, high-level, and widely used interfaces is the Python interface.

Flexible and changeable architecture: TensorFlow is designed to work on various hardware and network architectures. It is so abstract that the code used to train the same model is almost identical, whether running on a single machine or a computing cluster.

Product-oriented: TensorFlow has been developed by the Google brain team as a tool for large-scale development and serving machine learning models. Its design philosophy is to build the entire process from design to product.

TensorFlow is also a numerical computing library, and you can use its mathematical operation functions at will to give full play to the total computing power of the hardware. The system builds a data model on the platform. It predicts the evaluation value of the training process by training the neural network algorithm and finally obtains the recommendation of the training program.

Collaborative filtering based on LSTM recurrent neural network

In this study, not only should the patients’ basic information be considered, but also, during the training process, attention should be paid to the changes in the evaluation data and training methods in the time series. The latent semantic model of Collaborative Filtering can meet data analysis needs. Weighted Matrix Factorization is a recommendation algorithm for latent semantic optimization model, which can perform data analysis and processing in two dimensions basic patient information and auxiliary information. Then, according to the weight expression, the input elements of the LSTM model are obtained to train to predict the evaluation value of the patient training process and select the training program.

Basic information about patients: This information includes the patient's gender, age, cause of illness, training location, and other attribute information. The gender and age of the patient have a specific influence on the recovery time and effect on the patient. The cause of illness is essential information to determine the direction of rehabilitation training. According to different disease causes, choosing targeted training can significantly improve the Treatment effect and reduce the treatment cycle.

Assistance information of patients: The patient auxiliary information includes the training program selected in the time series and the motor function evaluation value during the training process of the recovered patients in the rehabilitation center. The normalization of training time and the acceleration of training speed show that the patient is recovering slowly, and the training program of the whole process is essential. The continuous adjustment of the training program based on the results of the assessment of motor function in a fixed period can ensure that each step of the patient's rehabilitation training is accurately arranged and also provides the efficiency and reliability of the patient's training process.

Recurrent Neural Networks can predict timing problems, and the time nodes in all hidden layers of RNN are connected according to time rules. Neurons in each hidden layer can record patient data at the current moment and use the neuron's output data at the last moment as input information. In the internal structure of a standard RNN, tanh() is the activation function, s_t is the state value of A, and it is also the output value Y_t at the previous moment. The specific numerical calculation is as formula (2):

$$s_t=\tanh (s_{t-1},X_t,)$$

(2)

If the given RNN input sequence is x = [X_1, X₂,…, X_t], at time t, the state of the RNN hidden layer is s_t, and the output value is Y_t, and formulas (3) and (4) is derived:

$$Y_t=\sigma (V\cdot s_t)$$

(3)

$$s_t=\mathrm{g}(U\cdot X_t+W\cdot s_{t-1})$$

(4)

In the above formula, V represents the weight matrix of the standard RNN output layer; U represents the weight matrix of the input layer; W represents the weight matrix of the model hidden layer state s_t−1 as the input at time t; σ represents the output layer activation function; g represents the hidden layer activation function. Substitute equation (4) into equation (2) cyclically, as shown in equation (5):

$$Y_t=V\cdot g(U\cdot X_t+W\cdot g(U\cdot X_{t-1}+W\cdot g(U\cdot X_{t-2}+\dots )))$$

(5)

From equation (4), it can be obtained that at time t, the output value Y_t of RNN will be affected by factors such as X_t、X_t−1、X_t−2、X_t−3、….

The advantage of RNN is that the parameters of the state transition function are the same. Specifically, the input layer, hidden layer, and output layer of RNN share parameters U, V, and W. The RNN cyclic neural learning and training algorithms all use the Backpropagation through time (BPTT), and its optimized loss function L is the transformed form of the RNN learning process, and sums the loss values of all time nodes of the patient's physiological indicators. The loss function during backpropagation is derived as follows, and the final loss value is obtained by summing the loss functions at various positions in the sequence:

$$L=\sum _{t-1}^T{L}_t$$

(6)

After that, the calculation of the RNN output layer and the hidden layer is formulated in the form of (6), where the f activation functions of the RNN output layer and the hidden layer are the soft max function and the tanh() function respectively, and the gradient calculation formula of the weight matrix of the output layer is as shown in formula (7).

$$y=f(Wx+b)$$

(7)

$${\displaystyle \frac{\mathrm{\partial }L}{\mathrm{\partial }V}=\sum _{t=1}^T{\displaystyle \frac{\mathrm{\partial }{L}_t}{\mathrm{\partial }V}=\sum _{t=1}^T{\displaystyle \frac{\mathrm{\partial }{L}_t}{\mathrm{\partial }{o}_t}{\displaystyle \frac{\mathrm{\partial }{o}_t}{\mathrm{\partial }V}=\sum _{t=1}^T({\hat{y}}_t-y_t)(h_t{)}^T}}}}$$

(8)

In the BPTT path, the gradient loss of the current time t and the next time t + 1 jointly determines the gradient loss of the sequence position, and the state gradient of the hidden layer of the current RNN time t is defined as a formula (8). Calculate the gradient h_t of the hidden layer of the RNN at the current moment; as shown in formula (9), the calculation formula for the gradient ${\delta }_T$ at the final moment T is shown in (10).

$${\delta }_t={\displaystyle \frac{\mathrm{\partial }L}{\mathrm{\partial }{h}_t}}$$

(9)

$${\delta }_t={\displaystyle \frac{\mathrm{\partial }{L}_t}{\mathrm{\partial }{o}_t}{\displaystyle \frac{\mathrm{\partial }{o}_t}{\mathrm{\partial }{h}_t}+{\displaystyle \frac{\mathrm{\partial }{L}_t}{\mathrm{\partial }{h}_{t+1}}{\displaystyle \frac{\mathrm{\partial }{h}_{t+1}}{\mathrm{\partial }{h}_t}=V^T(y_t-y_t)+W^T{\delta }_{t+1}diag(1-{(h_{t+1})}^2)}}}}$$

(10)

$${\delta }_T={\displaystyle \frac{\mathrm{\partial }{L}_T}{\mathrm{\partial }{o}_T}{\displaystyle \frac{\mathrm{\partial }{o}_T}{\mathrm{\partial }{h}_T}=V^T({\hat{y}}_T-y_T)}}$$

(11)

After obtaining the gradient of the hidden layer state at the current moment, formulas (11) and (12) are obtained to calculate the RNN hidden layer weight W and the input layer parameter U.

$${\displaystyle \frac{\mathrm{\partial }\mathrm{L}}{\mathrm{\partial }\mathrm{W}}=\sum _{\mathrm{t}=1}^{\mathrm{T}}{\displaystyle \frac{\mathrm{\partial }{\mathrm{L}}_{\mathrm{t}}}{\mathrm{\partial }{\mathrm{h}}_{\mathrm{t}}}{\displaystyle \frac{\mathrm{\partial }{\mathrm{h}}_{\mathrm{t}}}{\mathrm{\partial }\mathrm{W}}}}=\sum _{\mathrm{t}=1}^{\mathrm{T}}\mathrm{diag}(1-{({\mathrm{h}}_{\mathrm{t}})}^2){\delta }_{\mathrm{t}}({\mathrm{h}}_{\mathrm{t}-1}{)}^{\mathrm{T}}}$$

(12)

$${\displaystyle \frac{\mathrm{\partial }\mathrm{L}}{\mathrm{\partial }\mathrm{U}}=\sum _{\mathrm{t}=1}^{\mathrm{T}}{\displaystyle \frac{\mathrm{\partial }{\mathrm{L}}_{\mathrm{t}}}{\mathrm{\partial }{\mathrm{h}}_{\mathrm{t}}}{\displaystyle \frac{\mathrm{\partial }{\mathrm{h}}_{\mathrm{t}}}{\mathrm{\partial }\mathrm{U}}}}=\sum _{\mathrm{t}=1}^{\mathrm{T}}\mathrm{diag}(1-{({\mathrm{h}}_{\mathrm{t}})}^2){\delta }_{\mathrm{t}}({\mathrm{x}}_{\mathrm{t}}{)}^{\mathrm{T}}}$$

(13)

RNNs suffer from vanishing and exploding gradients when dealing with longer sequences. This limitation makes RNN unable to obtain long-distance information. In order to overcome this problem, the RNN model is improved to obtain the LSTM model, which solved the long-distance sequence dependence of RNN.

LSTM-related researchers added three gates within the neuron to filter the calculation results, the three gates are the input gate, the forget gate, and the output gate. Each time the network is run, the neurons are filtered using a forget gate, discarding inconsistent computations and keeping only useful data. The working principle of LSTM is “one in and two out.” The standard structure is a storage unit state and three gates (input gate i_t, forget gate f_t, and output gate o_t).

The unit derivation process of LSTM is: the network input at time t is: the input value is x_t, the output value h_t−1 and the gating unit c_t−1; The network output at time t is: the output value h_t and the gating unit c_t. In the LSTM network, the function of the forget gate is to determine the influence of c_t−1 on c_t, the function of the input gate is to determine the degree of influence of x_t on c_t, and the function of the output gate is to control the degree of influence of c_t on h_t. Equation (13), Equation (14), and Equation (15) are the formulas for forgetting gate, input gate, and output gate, respectively, as follows:

$${\mathrm{f}}_{\mathrm{t}}=\sigma ({\mathrm{W}}_{\mathrm{f}}\cdot {\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{W}}_{\mathrm{f}}\cdot {\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{f}})$$

(14)

$${\mathrm{i}}_{\mathrm{t}}=\sigma ({\mathrm{W}}_{\mathrm{i}}\cdot {\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{W}}_{\mathrm{i}}\cdot {\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{i}})$$

(15)

$${\mathrm{o}}_{\mathrm{t}}=\sigma ({\mathrm{W}}_{\mathrm{o}}\cdot {\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{W}}_{\mathrm{o}}\cdot {\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{o}})$$

(16)

Among them, f_t, i_t, and o_t are the calculation results of the three-gate states, respectively, W_f, W_i, and W_o are the weight matrices of the three-gates, respectively, and b_f, b_i, and b_o are the output bias terms of the three-gates, respectively. The final output of LSTM is shown in formula (18), there is an output gate, and the output of the forget gate is determined together, and the calculation process is shown in formula (16) and formula (17).

$${\mathrm{c}}_{\mathrm{t}}^{\sim }=\tanh ({\mathrm{W}}_{\mathrm{c}}\cdot {\mathrm{h}}_{\mathrm{t}-1}+{\mathrm{W}}_{\mathrm{c}}\cdot {\mathrm{x}}_{\mathrm{t}}+{\mathrm{b}}_{\mathrm{c}}^{)}$$

(17)

$${\mathrm{c}}_{\mathrm{t}}={\mathrm{f}}_{\mathrm{t}}\circ {\mathrm{c}}_{\mathrm{t}-1}+{\mathrm{i}}_{\mathrm{t}}\circ {\mathrm{c}}_{\mathrm{t}}^{\sim }$$

(18)

$${\mathrm{h}}_{\mathrm{t}}={\mathrm{o}}_{\mathrm{t}}\circ \tanh ({\mathrm{c}}_{\mathrm{t}})$$

(19)

In the above formula, ${\mathrm{c}}_{\mathrm{t}}^{\sim }$ is the input unit state at time t; $\circ$ is the element-wise multiplication; $\tanh$ is the activation function; ${\mathrm{b}}_{\mathrm{c}}$ is the input unit state bias term; and ${\mathrm{W}}_{\mathrm{c}}$ is the input unit state weight matrix. The training algorithm of LSTM is based on the back-propagation algorithm, which mainly modifies the weight matrix appearing in the neural network to minimize the error between the actual output and the ideal output.

Patient information and auxiliary information are descriptions of patient attributes and training process data, which belong to a kind of time series text structure data. After data classification and extraction of information features and preprocessing, digital sequence input in time series can be obtained and entered into LSTM. The training is carried out in the algorithm model, and the algorithm structure diagram is shown in Figure 1.

Figure 1.

Algorithm structure diagram.

Data collection

This system serves the field of rehabilitation medicine. The data sources are composed of basic patient information, electronic files, electronic medical records, doctor's order system, use records of rehabilitation equipment such as rehabilitation robots, training data, and physiological indicators of patient training. Among them, the patient's physiological indicators and training data need to be collected through the data collection module of the device.

The patient's training records, physiological data, and equipment usage records need to be transmitted through the communication interface of the rehabilitation equipment. The exoskeleton upper limb robot is taken as an example for analysis.

The basic parameters such as the motor and acceleration, peak speed, and movement direction of the robotic arm joint are selected in the file configuration of the host computer system. The host computer can accept control commands to run control strategies for different rehabilitation training programs. The control system controls the control system according to the host computer. The command selects the exercise mode, and the subroutine is called by sending the corresponding table value. From the time of accepting the designation to the end of the task, the main control unit Trio module has been monitoring the data command, and the rehabilitation information management system database system can also obtain the table value through data communication, sensor information.

Each type of rehabilitation robot has its host computer control system, which is mainly designed for information access and command. The information management system and the host computer control system can communicate through Ethernet technology. The upper limb rehabilitation robot also has a patient physiological indicator module, which mainly detects physiological indicators such as heart rate, blood pressure, blood oxygen saturation, and other physiological indicators of the patient during training.

In addition to the patient training records and physiological indicators collected through the robot's own module, the database can also continuously enrich some evaluation indicators during the patient's rehabilitation process through the manual input of the information by medical staff on the client, including blood glucose levels, blood oxygen saturation, muscle Strength, muscle tone, joint range of motion (refers to the angle of motion or rotation that a joint passes through during movement, and is an important indicator for evaluating joint function), motor function staging (Brunnstrom staging), Ueda assessment grade, etc.

The grading of muscle strength increases gradually from 0 to 5: 0 is complete paralysis; 1 has muscle contraction without slight movement; 2 can move the limbs slightly but cannot leave the bed; Leave the bed surface; Grade 4 can support the patient to overcome resistance, and the limbs have certain resistance; Grade 5 is normal muscle strength.

Muscle tone (Ashworth) is also divided into 0–5 grades: grade 0 is normal muscle tone; Grade 1 muscle tone is slightly increased, and the affected limb is suddenly stuck or released during passive flexion and extension; Grade 2 muscle tone is slightly increased, there is less resistance after joint movement; Grade 3 muscle tone increases significantly, but there is obvious resistance to the movement of the affected limb; Grade 4 muscle tone increases significantly, and passive movement is difficult; Grade 5 is rigid, and the affected limb cannot move when passively flexed or extended.

Motor function staging (Brunnstrom staging) divides the recovery process of limb movement into six stages, namely stage I without slight movement, stage II with slight flexion and joint movement, stage III with extension and joint movement, and stage IV with partial separation (Partial pronation and supination of the elbow joint), complete separation in stage V (90° abduction of elbow extension and 180° forward flexion), and normal stage VI. The 0, (1, 2), (3, 4, 5, 6), (7, 8), (9, 10, 11), and 12 grades of Ueda's 12-grade assessment method are more detailed for the staging of limb motor function division.

This system is an information management system for rehabilitation medicine. The data set is composed of valid data for limb motor function rehabilitation selected based on the clinical data set of the hospital rehabilitation center. As shown in Table 1, the dataset mainly includes a patient information table (PatientInformation.txt) and a patient auxiliary information table (PatientAuxiliaryInformatio.txt). The patient information table contains the basic information of the recovered patient's serial number, gender, age, and training location, and the patient auxiliary information table mainly contains the functional evaluation value, treatment method, training plan, and evaluation time during the rehabilitation process. The functional evaluation values (e.g., Ueda assessment) are clinically assessed by experienced rehabilitation physicians or therapists using standardized scales (e.g., Brunnstrom staging, Ueda 12-grade method) and then entered into the database.

Table 1.

Experimental dataset.

File name	PatientInformation.txt	PatientAuxiliaryInformatio.txt
File description	Patient information sheet	Patient assistance Information sheet
File size	2.3M	30.6M
Main content	Patient number Gender Age Training position	Ueda function evaluation value Treatment Training method Assessment time

Data preprocessing

Through the data collection link, the changes in key physiological indicators of patients within 1 month or 1 year can be obtained. The data set consists of the data set provided to the research institute by the Rehabilitation Department of a Hospital in China, including the data of patients whose physiological indicators have returned to the normal range after treatment for more than one year. The patient auxiliary information table is modeled by serialization to generate the latent representation vector of physiological data. The purpose of the model is to learn according to the patient attribute parameters combined with the weight parameters of the model to form input neurons.

Preprocessing is to form a sequence of patient information and patient auxiliary information into a digital sequence according to time and change the information in the original document to form a corresponding digital information table, converting the words that appearing in the information into a sequence of numbers. In this process, some constraints are added to control the quality of information to improve the data's validity.

Gender: The gender of the patient is 00 for male and 01 for female.

Age: Patient age grades are converted into numeric sequences. Four grades were set up, those under 45 years old (00), 45 years old-55 years old (01), 56 years old-65 years old (02), and over 65 years old (03), and the patients were divided into four categories, and the data were sorted.

Training position: Upper limb rehabilitation training (00) and Lower limb rehabilitation training (01). The training parts of the patients are different, and the rehabilitation equipment used is also different. The patients selected in this experiment are all patients who have used similar upper and lower limb rehabilitation robots for motor function training, and the rehabilitation equipment defaults to the upper and lower limb rehabilitation robot equipment.

Cause of illness: Different rehabilitation patients have different reasons for the initial limb injury, and the causes of illness are classified according to the different nature of limb injury. Systemic diseases (00): such as Japanese encephalitis, cardiovascular and cerebrovascular diseases, poliomyelitis, etc.; accidental injuries (01): such as traffic accidents, sports injuries caused by muscle and motor function damage, etc.; physical factors (02): such as burns, falls, etc.; chemical factors (03): such as drugs, alcoholism, etc.; genetic factors (04): deformity, developmental delay, etc.; others (05): such as congenital physical disabilities, etc.

The most important evaluation index in the patient's motor function training process is the Ueda function evaluation form. Since Ueda is more accurate than Brunnstrom's functional staging, the Ueda function evaluation form is clinically used as the evaluation standard for patient training projects. The clinical data sample table that describes the rating scale is divided into 13 grades.

The data includes not only the grade evaluation during the training process of the patient, but also the treatment method and training method of the patient. During the treatment process, the evaluation results are recorded every half month, and the training method is adjusted.

Treatment method: Active training (00), Passive training (01).

Training method: 5r/min, within half an hour (00); 5r/min, between half an hour and an hour (01); 10r/min, within half an hour (02); 10r/min, between half an hour and an hour (03); 15r/min, within half an hour (04); 15r/min, between half an hour and an hour (05); 20r/min, within half an hour (06); 20r/min, between half an hour and an hour (07).

After preprocessing patient information and auxiliary information to obtain a digital sequence, it cannot be directly used as an input for training the model. It is necessary to convert the digital sequence into a word vector sequence, that is, each number is represented by a fixed-dimensional vector. The sequence is vectorized with a word vector sequence a of length n $\{{\alpha }_1,{\alpha }_2,\dots ,{\alpha }_{\mathrm{n}}\}$ , as the LSTM layer.

Experiment analysis

The primary function of the LSTM layer is to train the data set, train the patient's functional evaluation value and training method, and finally provide the patient with the predicted functional evaluation value and training method that conforms to their basic information. The training plan is composed of the training method of each node. The LSTM model is trained by minimizing the Mean Squared Error (MSE) between the predicted functional evaluation values and the actual clinical assessment values.

After preprocessing and vectorization, each patient's input feature vector at time step t contains 8 dimensions, namely: gender encoding, age group encoding, training site encoding, etiology encoding, treatment method encoding, training method encoding, current Ueda evaluation value, evaluation time step. For example, a male patient aged 45–55 undergoing upper limb rehabilitation due to stroke (classified as a ‘systemic disease’), at a certain evaluation node, the input can be represented as [00, 01, 00, 00, 00, 02, 7, 3]. The output vector of the model is the predicted Ueda evaluation value (a continuous value or classification value of 0–12) for the next evaluation node of the patient, which is the label used during training. In the end, the output of a series of time steps constitutes a complete training scheme sequence.

This experiment collected real clinical data from 50 patients. After denoising and screening, a total of 50,000 valid time-series records were obtained, covering the entire process from treatment to rehabilitation and discharge. Among them, 36,000 are used as the training set, 7000 are used as the validation set, and 7000 are used as the test set. In order to improve the generalization ability of the model under limited data, we adopted data augmentation techniques (such as small scaling and translation of time-series data) to expand the training dataset, rather than using transfer learning. The prediction accuracy of the training plan and rehabilitation time is obtained after analyzing the output of the model, as shown in Figure 2. According to the accuracy shown in Figure 2, it is determined whether the training plan and recovery time are more accurate. As shown in Figure 2, the model's accuracy on the training and validation sets improves rapidly during the initial training epochs and tends to converge after approximately 25 epochs, with minor fluctuations remaining within a certain range. Therefore, the training plan recommended model in the system has certain data support. In this paper, the rehabilitation training program is analyzed and recommended to patient users and rehabilitation therapists through the model predicted by big data to help patients understand the approximate training program and time and to provide therapists with a reference recommended program.

Figure 2.

Accuracy of the model.

In the two-layer LSTM model, the hidden layer dimension of the first layer LSTM is set to 50, and the hidden layer dimension of the second layer LSTM is set to 30. This dimension is consistent with the dimension of the input word vector and is used to capture higher-order temporal features. After the second layer of LSTM, connect a fully connected layer and map the output of the hidden layer with dimension 30 to the predicted value of 1D (i.e., the predicted Ueda function evaluation value). The final output layer designed in this way is consistent with the predicted target dimension.

As shown in Figure 2, the model's accuracy on the training and validation sets improves rapidly during the initial training epochs and tends to converge after approximately 25 epochs, with minor fluctuations remaining within a certain range.

This study is a recommendation about the patient's rehabilitation training program. In the process of algorithm implementation, it is necessary to predict the patient's Ueda function evaluation value. The accuracy of the algorithm can be obtained by comparing the test set with the predicted evaluation value. There are mainly two evaluation indicators. For analysis, mean absolute error (MAE) and root-mean-square error (RMSE), as shown in Equations (19) and (20).

$$\mathrm{MAE}={\displaystyle \frac{1}{\mathrm{N}}\sum _{\mathrm{i}.\mathrm{j}}{\mathrm{Z}}_{\mathrm{i},\mathrm{j}}^{\mathrm{P}}|{\mathrm{r}}_{\mathrm{ij}}-{\hat{\mathrm{r}}}_{\mathrm{ij}}|}$$

(20)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{N}}\sum _{\mathrm{i},\mathrm{j}}{\mathrm{Z}}_{\mathrm{i},\mathrm{j}}^{\mathrm{P}}{({\mathrm{r}}_{\mathrm{ij}}-{\hat{\mathrm{r}}}_{\mathrm{ij}})}^2}}$$

(21)

Specifically, i indexes the sample, j indexes the evaluation metric (here j = 1 for Ueda score), r_ij and ${\hat{r}}_{ij}$ are the actual and predicted values, and $Z_{i,j}^P$ is a binary indicator for test set membership.

Here, N and Z respectively represent the total number of test sets and the binary matrix of the test evaluation value. MAE and RMSE can represent the error relationship between the actual and predicted evaluation values. The effect of RMSE is more obvious and can be more intuitive.

Accuracy of performance predictions

The smaller the RMSE value, the higher the prediction accuracy. During the model training process, the change process of the RMSE values of the training set and the test set is shown in Figure 3. The changes of the RMSE indicators of the verification data set and the test data set are relatively similar. As the number of times increases, the RMSE value between the two continues to converge, and the performance index changes more significantly.

Figure 3.

Model training process.

Based on the predicted evaluation value and actual value of the training process output by the model in this study, according to formula (20), the RMSE value is 0.7531, which is small, indicating that the predicted evaluation value is more accurate.

Using the same data set, the RNN cyclic neural network algorithm is used to judge the accuracy of the predicted evaluation value. Based on the evaluation value and the predicted evaluation value, the RSME value is calculated to be 1.0878, which is relatively high. It can be seen that the LSTM-based collaborative filtering algorithm has a higher accuracy rate than the RNN algorithm, and the accuracy rate is greatly improved.

Conclusions

With the continuous deepening of medical informatization, management systems applied in the medical field emerge in endlessly, the rehabilitation medical industry is also developing rapidly, and the rehabilitation centers of hospitals are gradually entering the transformation stage. In recent years, the continuous development and introduction of rehabilitation robots, artificial intelligence and ar/vr technology have strengthened the management of basic information of rehabilitation institutions; In view of the scattered and unsystematic clinical use data in the development of rehabilitation industry, the information management system of rehabilitation center needs to constantly innovate. This paper focuses on the recommendation technology of rehabilitation training scheme in the rehabilitation management system based on artificial intelligence.

The research of this paper may have the following deficiencies. The data set used in this system is the data provided by the rehabilitation department of a domestic hospital. Firstly, the data is incomplete, and there is no systematic data classification; Secondly, patients who use rehabilitation equipment have fewer clinical cases. If the public platform can provide more accurate and comprehensive medical rehabilitation data, it will greatly improve the data comprehensiveness of the model and the reliability of the analysis results. In this study, LSTM recurrent neural network combined with collaborative filtering algorithm is used to establish the recommendation model of rehabilitation training scheme. In the next step, attention mechanism can be added to the recommendation algorithm or combined with other traditional recommendation models, and CT image recognition and other technologies can be added to realize more accurate recommendation of rehabilitation training scheme.