CN-122025003-A - Multi-parameter self-adaption-based child personalized vision training system and method

CN122025003ACN 122025003 ACN122025003 ACN 122025003ACN-122025003-A

Abstract

The invention belongs to the technical field of vision and vision training, and particularly relates to a multi-parameter self-adaption-based personalized vision training system and method for children, comprising the following steps of S1, acquiring clinical baseline data of a user, wherein the information at least comprises basic physiological parameters and initial vision ability evaluation targets of the user; and S2, inputting the clinical baseline data and the multi-mode dynamic response data into a fusion analysis model constructed based on an information bottleneck theory. The invention can realize personalized, dynamic, safe and controllable visual function rehabilitation training in the true sense through multi-mode data fusion analysis, user modeling based on an information bottleneck theory, a double-time-scale self-adaptive decision engine and a man-machine collaborative calibration mechanism, and ensure the scientificity and effectiveness of a training scheme through a continuous learning mechanism.

Inventors

WANG JIANGBO
WANG JIAO
QIU XIAOJUAN
WANG CHANGZAI

Assignees

杭州深康明视科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. The multi-parameter self-adaptive child personalized vision training method is characterized by comprising the following steps of: S1, acquiring clinical baseline data of a user, wherein the information at least comprises basic physiological parameters and initial visual ability evaluation targets of the user; S2, inputting the clinical baseline data and the multi-mode dynamic response data into a fusion analysis model which is built; The method comprises the steps of obtaining a joint representation vector for encoding a visual state of a user through characterization learning of a model, and generating a comprehensive evaluation index for quantitatively characterizing the current visual state and the capability level of the user based on vector decoding; s3, according to the comprehensive evaluation index, automatically matching and generating an initial personalized training program from a preset multidimensional visual training task library, and pushing the program to user terminal equipment; And S4, acquiring task performance data and non-contact physiological behavior data of a user in real time in the process of executing a training task at the terminal, and dynamically optimizing at least one parameter of a subsequent training task by utilizing an adaptive decision engine based on the task performance data and the non-contact physiological behavior data so as to update the training plan.
2. The method according to claim 1, wherein in step S1, the user' S basic physiological parameters include one or more of age, refractive status, and interpupillary distance; The initial visual ability assessment objective comprises one or more of eyesight, contrast sensitivity, stereo vision sharpness, fusion range, and binocular inhibition degree; The multi-mode dynamic response data comprises time sequence eye movement track data acquired by the eye movement tracking equipment, task performance data recorded by the interactive system and implicit behavior preference data obtained by analyzing a user operation sequence.
3. The method according to claim 1, wherein in step S2, the fusion analysis model is trained by minimizing the following information bottleneck loss function: wherein X is input multi-mode dynamic response data, Y is a visual ability evaluation target, Z is a joint representation vector obtained by model learning, Representing mutual information between the joint representation vector Z and the original input data X, And representing mutual information between the joint representation vector Z and the target variable Y, wherein beta is a super parameter for controlling the compression degree, and the comprehensive evaluation index is obtained by decoding the joint representation vector Z.
4. The method according to claim 1, wherein in step S2, the fusion analysis model employs a cross-attention-based multi-modal adaptive weighted fusion mechanism prior to generating the joint representation vector: Firstly, mapping dynamic response data of different modes to the same feature space through independent encoder networks; Secondly, taking the sequential eye movement track data characteristics as query vectors, taking the data characteristics of other modes as key vectors and value vectors, and dynamically generating attention weights for different modes through cross attention calculation; And finally, carrying out weighted fusion on the multi-mode features according to the attention weight, and taking the fused features as the input of an information bottleneck layer to generate a joint representation vector which can reflect the real-time state of the user and the multi-mode association.
5. The method according to claim 1, wherein in step S4, the optimization process of the adaptive decision engine comprises two time scale adjustments: Based on continuous task performance data acquired in a preset sliding time window, and combining with real-time estimated user concentration and fatigue index, dynamically adjusting at least one difficulty parameter, presentation duration or feedback form of a subsequent single task or task batch in a current training unit; After completing a preset training period, acquiring the performance data of a user on two tasks by alternately presenting a comparison test link of a suggested advanced difficulty task and a reference difficulty task; And carrying out significance difference analysis on the performance data based on statistical hypothesis test, and deciding whether to formally update the global training difficulty parameter in a subsequent training period according to analysis results.
6. The method according to claim 1, wherein in the step S4, the non-contact physiological behavior data includes facial expression characteristics, head posture changes and transient frequency of the user acquired by the terminal camera and analyzed by the computer vision algorithm, and the non-contact physiological behavior data are used for assisting in estimating concentration, cognitive load and training comfort of the user.
7. The method according to claim 1, further comprising step S5 of periodically generating and pushing a visual training effect report to a guiding physician terminal, wherein the effect report comprises a user capability evolution trend, a model decision basis and a period verification result; The self-adaptive decision engine adopts inverse reinforcement learning, and utilizes the feedback sample to continuously optimize a strategy model in the engine, so that the system decision is gradually attached to the professional judgment logic of a clinician.
8. A multi-parameter adaptive child personalized visual training system, which is characterized in that the method for realizing any one of the claims 1 to 7 comprises the following steps: The user terminal module is configured with a display unit, an interaction unit and an eye movement tracking unit and is used for presenting training tasks and collecting interaction and eye movement data; The data preprocessing module is used for cleaning, synchronizing and primarily extracting features of the collected original multi-mode data; the multi-mode fusion and user modeling module is used for running the fusion analysis model to generate comprehensive evaluation indexes and state representations of the user; The personalized plan generation and safety calibration module is used for matching training tasks according to user states and generating an initial plan by applying safety rules; The self-adaptive decision engine module is used for dynamically optimizing a subsequent training plan according to the real-time training data; The task library management module is used for storing and managing multidimensional visual training task materials and generating logic; And the doctor cooperation platform module is used for displaying the effect report and the user model state and receiving a calibration instruction of a doctor.
9. An electronic device comprising a memory for storing processor-executable instructions, wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any of claims 1-7.
10. A computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1 to 7.

Description

Multi-parameter self-adaption-based child personalized vision training system and method Technical Field The invention belongs to the technical field of vision and vision training, and particularly relates to a multi-parameter self-adaption based personalized vision training system and method for children. Background Children with dysplasia of visual functions such as amblyopia and strabismus usually need to be intervened through scientific visual training. Currently, computer-aided vision training systems have become increasingly popular. However, the prior art schemes have obvious limitations that firstly, most training schemes of the system are set on the basis of initial evaluation in a 'one-time' manner, and lack of a closed-loop mechanism for dynamic adjustment according to the real-time capability state (such as fatigue degree, attention fluctuation and learning progress) of a user, so that the training efficiency is low, secondly, the self-adaptive decision of the system is fully automatic, professional supervision and calibration of a clinician are lacked, the risk of invalid training or discomfort of the user caused by improper setting of difficulty exists in a training scene of a court, and finally, the system cannot continuously learn from the intervention of the doctor, and the decision logic of the system is difficult to align with clinical best practice. For example, some adaptive training schemes are disclosed in the prior patent literature that generally adjust task difficulty based on user eye movement trajectories or test achievements. However, these schemes focus on a single automated adjustment and do not build a collaborative mechanism that incorporates clinical expert experience into the decision-making closed loop and enables system self-iteration. Meanwhile, for efficiently extracting the core state characteristics of the user from multi-mode data (eye movement, behaviors and expressions), the lack of an advanced modeling method leads to inaccurate evaluation, thereby affecting the matching degree of the personalized plan. Disclosure of Invention The invention aims to provide a multi-parameter self-adaptive child personalized visual training system and a multi-parameter self-adaptive child personalized visual training method, which can realize personalized, dynamic, safe and controllable visual function rehabilitation training in a real sense through multi-mode data fusion analysis, user modeling based on an information bottleneck theory, a double-time-scale self-adaptive decision engine and a man-machine collaborative calibration mechanism, and ensure the scientificity and effectiveness of a training scheme through a continuous learning mechanism. The technical scheme adopted by the invention is as follows: a multi-parameter self-adaption based personalized visual training method for children comprises the following steps: S1, acquiring clinical baseline data of a user, wherein the information at least comprises basic physiological parameters and initial visual ability evaluation targets of the user; S2, inputting the clinical baseline data and the multi-mode dynamic response data into a fusion analysis model constructed based on an information bottleneck theory, and obtaining a joint representation vector for compression characterization of the current core visual state of a user through learning by optimizing an information bottleneck loss function; S3, matching and generating an initial personalized training program from a preset multidimensional visual training task library according to the comprehensive evaluation index of the user, wherein the initial task difficulty in the program is calibrated through a safety constraint module so as to be set in a range of 70-80% of an estimated threshold value based on the user capacity; S4, collecting task performance data and non-contact physiological behavior data of a user in real time in the process of executing a training task by the user in a terminal, and dynamically optimizing at least one difficulty parameter, task type or visual presentation characteristic of a subsequent training task by utilizing an adaptive decision engine so as to update a training plan; S5, periodically generating a visual training effect report and pushing the visual training effect report to a terminal for guiding a doctor, wherein the effect report comprises a user capacity evolution trend, a model decision basis and a period verification result; the self-adaptive decision engine adopts inverse reinforcement learning, and utilizes the feedback sample to continuously optimize a strategy model in the engine, so that the system decision is gradually attached to the professional judgment logic of a clinician. In the step S1, the user basic physiological parameters include one or more of age, refractive state, and interpupillary distance; The initial visual ability assessment objective comprises one or more of eyesight, contrast sensitivity, stereo vision sh