CN-121983234-A - Training method for generating AI model by personalized lower limb rehabilitation training scheme and using method

CN121983234ACN 121983234 ACN121983234 ACN 121983234ACN-121983234-A

Abstract

The invention discloses an AI model training method generated by a personalized lower limb rehabilitation training scheme and a using method thereof. The training method comprises the steps of S1, obtaining a training sample set containing rehabilitation data of a plurality of historical patients, carrying out data enhancement processing on the training sample set to generate a virtual sample, S2, training basic information of the patients in the training sample set by adopting a decision tree algorithm to obtain a feature classification model, wherein the feature classification model is used for mapping the basic information of the patients into feature classification information, S3, modeling a formulating and adjusting problem of a rehabilitation training scheme as a Markov decision process based on the extended training sample set and the feature classification information, S4, training to obtain a depth Q network model by adopting a depth Q network algorithm based on the extended training sample set and the Markov decision process established in the S3, and forming a personalized lower limb rehabilitation training scheme by adopting the feature classification model and the depth Q network model to generate an AI model.

Inventors

JIANG XIONG
CHEN JUN
JIN FEI

Assignees

遵义市中医院

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (10)

1. The AI model training method for generating the personalized lower limb rehabilitation training scheme is characterized by comprising the following steps of: step S1, a training sample set is obtained, wherein the training sample set comprises rehabilitation data of a plurality of historical patients, and the rehabilitation data of each historical patient comprises patient basic information, training records and rehabilitation results; Step S2, training basic information of a patient in a training sample set by adopting a decision tree algorithm to obtain a feature classification model, wherein the feature classification model is used for mapping the basic information of the patient into feature classification information, and the feature classification information comprises a rehabilitation stage category, a training intensity level, a rehabilitation strategy type and a risk level; step S3, modeling the establishment and adjustment problems of the rehabilitation training scheme as a Markov decision process based on the expanded training sample set and the feature classification information; step S4, training to obtain a deep Q network model by adopting a deep Q network algorithm based on the expanded training sample set and the Markov decision process established in the step S3; The characteristic classification model and the depth Q network model form a personalized lower limb rehabilitation training scheme to generate an AI model, and training scheme parameters output by the AI model comprise a movement speed, a movement angle range, training duration and a power assisting grade.
2. The personalized lower limb rehabilitation training protocol generation AI model training method of claim 1, wherein: In step S1: The patient base information includes age, time of onset, BMI, grade of muscle strength of the affected side, degree of joint movement of the affected side, improved Ashworth cramp score, brunstrom index, hemiplegic walking ability, lindmark index, pain vision simulation score, resting heart rate and subjective fatigue score; The training record comprises a training scheme adopted in a rehabilitation training period and corresponding training execution conditions, wherein training scheme parameters comprise a movement speed, a movement angle range, training duration and a power-assisted level, and the training execution conditions comprise training days actually completed and average heart rate monitoring data in a training process; the rehabilitation result comprises a muscle strength grade change value, a joint activity change value, an improved Ashworth spasm score change value, a Brunnstrom index change value, a hemiplegic walking ability change value, a Lindmark index change value and a pain vision simulation score change value before and after training; in step S2: the rehabilitation stage category comprises acute stage, early recovery stage, middle recovery stage and late recovery stage; The training intensity level includes a low intensity, a medium intensity, and a high intensity; the rehabilitation strategy type comprises passive training as a main part, active and passive combination and active training as a main part; The risk level includes low risk, medium risk and high risk.
3. The method for generating AI model training for personalized lower limb rehabilitation training scheme of claim 2, wherein at least part of the historical patient training records in the training sample set are derived from a lower limb rehabilitation training system, the lower limb rehabilitation training system comprises a lower limb rehabilitation training device, a detection unit, a PLC and a control system; The lower limb rehabilitation training device comprises a thigh fixing module, a shank fixing module, a base, a pedal, a sliding block, a sliding rail, an automatic reciprocating cylinder, a first hinge component, a second hinge component, a third hinge component and a fourth hinge component, wherein the thigh fixing module and the shank fixing module are connected through the first hinge component, the thigh fixing module is connected with the base through the second hinge component, the shank fixing module is connected with the sliding block through the third hinge component, and the sliding block is matched with the sliding rail; The detection unit comprises a first hinging component, a second hinging component, a pressure sensor array embedded with a pedal, a proportional pressure sensor arranged on an automatic reciprocating cylinder air supply pipeline, a travel sensor arranged on a piston rod of the automatic reciprocating cylinder, a heart rate monitoring bracelet worn on the wrist of a patient and a pain alarm button arranged on a base; the PLC is used for transmitting the data of the detection unit to the control system, receiving the instruction output by the control system and controlling the automatic reciprocating cylinder to execute the instruction; The control system comprises a data acquisition module, a parameter conversion module, a communication module and a man-machine interaction interface.
4. The method for generating AI model training for personalized lower limb rehabilitation training scheme as defined in claim 2, wherein in step S2, the training process of the feature classification model comprises: Patient basic information is used as input characteristics; The characteristic classification information is used as an output label, and the output label is obtained by manually marking the characteristic classification information of each historical patient according to the patient basic information, the training record and the rehabilitation result of the patient in the training sample set; And constructing a decision tree by adopting a C4.5 algorithm or a CART algorithm, selecting optimal split characteristics through information gain or genie unrepeace, and training to obtain the characteristic classification model.
5. The method for generating AI model training for personalized lower limb rehabilitation training protocol of claim 4, wherein in step S3, the modeling of the Markov decision process comprises: The state space S is defined as S= { patient basic information, rehabilitation stage category, training intensity level, rehabilitation strategy type, risk level, current training scheme parameter, trained days, accumulated rehabilitation progress, average heart rate of the last training period, average subjective fatigue degree score of the last training period }, wherein the accumulated rehabilitation progress comprises an reached muscle strength level change value, a joint activity change value, an improved Ashworth spasm score change value, a Brunnstrom index change value, a hemiplegic walking ability change value, a Lindmark index change value and a pain vision simulation score change value; the motion space A is defined as that one of the motion speed, the motion angle range, the training time length and the assistance level is increased, reduced or kept unchanged, the adjustment amplitude of each parameter is determined according to the training intensity level, and the adjusted parameter value is required to meet the safety constraint corresponding to the risk level and the preset parameter change rate constraint; The state transition probability P (s '|s, a) represents the probability of transition to the state s' after the action a is executed under the state s, and is obtained through actual transition statistics in the expanded training sample set; The reward function R (s, a, s') is calculated from rehabilitation progress, training completion rate, safety and physiological fitness over a period of time after the current adjustment, the rehabilitation progress being obtained by weighted summation of a muscle strength grade change value, a joint activity change value, an improved Ashworth cramp score change value, a brunstrom index change value, a hemiplegic walking ability change value, a Lindmark index change value and a pain vision simulation score change value.
6. The method for generating AI model training for a personalized lower limb rehabilitation training protocol as recited in claim 5, wherein the calculation formula of the reward function R (s, a, s') is: R=w1×rehabilitation progress score +: w2 x training completion rate score +w4×physiological fitness score-w3×security penalty; Wherein, rehabilitation progress score = α1 x muscle strength grade change value + α2 x joint mobility change value- α3 x improved Ashworth cramp score change value + α5 x Brunnstrom index change value + α6 x hemiplegic walking ability change value + α7 x Lindmark index change value- α4 x pain vision simulation score change value; the physiological adaptability score is calculated according to the average heart rate monitoring data and the change trend of the subjective fatigue degree score in the training period, and if the average heart rate monitoring data is in an effective target heart rate interval and the fatigue degree score does not exceed a preset threshold value, a forward score is obtained; The safety penalty is triggered when the rehabilitation progress is negatively changed, the pain score is increased or the training parameter change rate exceeds the smoothness constraint threshold, and w1, w2, w3, w4, alpha 1, alpha 2, alpha 3, alpha 4, alpha 5, alpha 6 and alpha 7 are preset weight coefficients.
7. The method for generating AI model training for a personalized lower limb rehabilitation training scheme as recited in claim 6, wherein in step S3, the safety constraint of the action space A comprises setting a range of values of training scheme parameters according to a risk level, determining an adjustment range according to a training intensity level, and setting an upper limit of a parameter change rate between two adjacent adjustments: Under the low risk level, the movement speed is limited within the range of 10-60 times/min, the movement angle is limited within the range of 30-120 degrees, the training time is limited within the range of 10-60 minutes, and the assistance level is limited within the range of 0-100%; Under the risk level, the movement speed is limited within the range of 10-40 times per minute, the movement angle is limited within the range of 30-90 degrees, the training duration is limited within the range of 10-40 minutes, and the assistance level is limited within the range of 30-100%; Under the high risk level, the movement speed is limited within the range of 10-20 times/min, the movement angle is limited within the range of 30-60 degrees, the training time is limited within the range of 10-20 minutes, and the assistance level is limited within the range of 50-100%; the upper limit of the parameter rate of change is used to limit the span of a single adjustment to prevent abrupt parameter changes.
8. The method for generating AI model training for a personalized lower limb rehabilitation training protocol of claim 7, wherein in step S4, said deep Q network comprises: The input layer receives a state vector, wherein the state vector comprises a numerical representation of basic information of a patient, independent heat codes of a rehabilitation stage class, independent heat codes of a training intensity level, independent heat codes of a rehabilitation strategy class, independent heat codes of a risk level, current training scheme parameters, trained days, accumulated rehabilitation progress, average heart rate and average subjective fatigue degree scores; A plurality of fully connected hidden layers, employing a ReLU activation function and a Dropout layer to prevent overfitting; The output layer outputs the Q value corresponding to each executable action; training is carried out through an experience playback mechanism and a target network updating strategy, and a loss function adopts a mean square error of a time sequence differential error.
9. Use of a personalized lower limb rehabilitation training scheme to generate AI models, characterized in that the AI models are trained by the method according to one of claims 1-8, the use comprising the steps of: Step U1, acquiring patient basic information of a new patient; Step U2, inputting the basic information of the patient into the feature classification model obtained by training in the step S2, and obtaining the feature classification information of the patient; step U3, constructing an initial state vector of the patient, wherein the current training scheme parameters adopt preset default initial values, the number of trained days is 0, and the accumulated rehabilitation progress is 0; step U4, inputting the initial state vector into the depth Q network model obtained by training in the step S4, and outputting the optimal action; and step U5, adjusting a default initial value according to the optimal action to obtain initial training scheme parameters.
10. The method for generating AI models for use with a personalized lower limb rehabilitation training protocol according to claim 9, further comprising the step of dynamically adjusting the training protocol: Step U6, training a new patient according to the initial training scheme parameters output in the step U5, periodically acquiring the current state information of the patient and generating a state vector; step U7, inputting the state vector obtained in the step U6 into the depth Q network model obtained in the training in the step S4, and outputting an optimal adjustment action; And step U8, updating the training scheme parameters of the patient according to the optimal adjustment action output by the step U7, and returning to the step U6 to continue execution until the rehabilitation training is completed.

Description

Training method for generating AI model by personalized lower limb rehabilitation training scheme and using method Technical Field The invention relates to the technical field of AI models, in particular to an AI model training method and a using method for generating an individualized lower limb rehabilitation training scheme. Background Lower limb dysfunction is a common sequelae of diseases such as cerebral apoplexy, spinal cord injury, bone surgery and the like, and seriously affects the daily life capacity of patients. Lower limb rehabilitation training is an important treatment means for recovering walking ability of patients. At present, the establishment of a lower limb rehabilitation training scheme mainly depends on clinical experience of rehabilitation doctors, and has the following technical problems: first, the degree of standardization is insufficient. Different rehabilitation doctors may develop training schemes with larger difference for patients with the same illness state, and the training effect lacks consistency guarantee. Second, personalized adjustment is difficult. The rehabilitation process of patients is affected by age, constitution, injury degree and other factors, and the individual differences are obvious. The traditional method is difficult to comprehensively consider all influencing factors, and cannot provide an accurate personalized scheme for each patient. Third, dynamic optimization is weak. Rehabilitation training is a dynamic process which lasts for weeks or even months, and the physical state of a patient is changed continuously. The existing method mainly depends on manual adjustment scheme after periodic evaluation by rehabilitation doctors, has low adjustment frequency and response lag, and is difficult to track the rehabilitation progress of patients in real time and optimize training parameters in time. Fourth, security risk control is inadequate. Improper training intensity may lead to secondary injury or arrest of rehabilitation progress, and existing methods lack systematic risk assessment and safety restraint mechanisms. In recent years, artificial intelligence technology has been widely used in the medical field. However, the existing rehabilitation training scheme generation method mostly adopts a simple rule matching or a single machine learning model, and cannot effectively solve the technical problems. The rule matching method relies on preset fixed rules and lacks learning capability, a single machine learning model can only give a static scheme according to the initial state of a patient and cannot realize dynamic adjustment in the training process, and the reinforcement learning method has dynamic optimization capability, but has the problems of high state space dimension, low sample utilization rate, difficult safety guarantee and the like in a rehabilitation training scene. Disclosure of Invention Aiming at the technical problems in the prior art, the invention provides an AI model generating training method for a personalized lower limb rehabilitation training scheme and a using method thereof, and the personalized training scheme generation and dynamic optimization based on individual characteristics of patients can be realized by combining decision trees and deep reinforcement learning. The technical scheme adopted by the invention is as follows: the AI model training method for generating the personalized lower limb rehabilitation training scheme comprises the following steps: step S1, a training sample set is obtained, wherein the training sample set comprises rehabilitation data of a plurality of historical patients, and the rehabilitation data of each historical patient comprises patient basic information, training records and rehabilitation results; Step S2, training basic information of a patient in a training sample set by adopting a decision tree algorithm to obtain a feature classification model, wherein the feature classification model is used for mapping the basic information of the patient into feature classification information, and the feature classification information comprises a rehabilitation stage category, a training intensity level, a rehabilitation strategy type and a risk level; step S3, modeling the establishment and adjustment problems of the rehabilitation training scheme as a Markov decision process based on the expanded training sample set and the feature classification information; step S4, training to obtain a deep Q network model by adopting a deep Q network algorithm based on the expanded training sample set and the Markov decision process established in the step S3; The characteristic classification model and the depth Q network model form a personalized lower limb rehabilitation training scheme to generate an AI model, and training scheme parameters output by the AI model comprise a movement speed, a movement angle range, training duration and a power assisting grade. Use of a personalized lower limb rehabilitation