Search

CN-122025077-A - Abnormal behavior identification and data quality assessment method in vestibular function examination

CN122025077ACN 122025077 ACN122025077 ACN 122025077ACN-122025077-A

Abstract

A method for identifying abnormal behavior and evaluating data quality in vestibular function examination. The invention discloses a method for identifying abnormal behaviors and evaluating data quality in vestibular function examination. The method comprises the steps of synchronously collecting eye movement data and head movement data of a detected person, extracting multidimensional fusion characteristics comprising eye closure duration, blink frequency, angular velocity RMS and static state proportion, identifying abnormal behaviors such as eye closure interference, head shaking or distraction by a rule engine or a machine learning classifier based on the characteristics, dynamically calculating data quality scores of 0-100 points by combining abnormal categories, confidence and continuity penalties, and carrying out visual marking and operation prompt on original eye movement records according to scoring grades (more than or equal to 90 are high in quality, 70-89 are available and <70 is unreliable). The invention realizes the semantic recognition of abnormal behaviors and the quantitative evaluation of data reliability, and remarkably improves the standardization level and clinical diagnosis efficiency of vestibular examination.

Inventors

  • XIAO WEIMIN
  • WANG JUN
  • ZHANG RUIGUANG

Assignees

  • 西安宏毅科技有限公司

Dates

Publication Date
20260512
Application Date
20251225

Claims (10)

  1. 1. The abnormal behavior identification and data quality assessment method in vestibular function examination is characterized by comprising the following steps of synchronously collecting eye movement data and head movement data of a subject in the vestibular function examination process, wherein the eye movement data are collected at a frame rate of not lower than 60Hz through an infrared video eye movement instrument, the head movement data are collected at a sampling rate of not lower than 100Hz through an inertial measurement unit IMU, and the two paths of data are aligned by a hardware trigger signal or a precision time protocol, and the synchronous error is not more than 5ms; Extracting at least one head motion feature from the head motion data, wherein the head motion feature comprises at least one of root mean square value RMS, euler angle change rate, resting state duty ratio and head shaking entropy of triaxial angular velocity, wherein the resting state duty ratio is defined as the time duty ratio that the angular velocity is less than 1 DEG/S and the acceleration vector mode approaches gravity acceleration g; Fusing the eye movement characteristics and the head movement characteristics to form a fused characteristic vector; Based on the fusion feature vector, identifying abnormal behavior categories of the detected person through a preset rule engine or a pre-trained machine learning classifier, wherein the abnormal behavior categories comprise at least one of eye closure interference, head shaking, distraction and normal coordination; Calculating the data quality score of the current inspection fragment according to the abnormal behavior category, the corresponding confidence coefficient and a preset deduction rule, wherein when the machine learning classifier is adopted, the confidence coefficient is the maximum category probability output by a model, when the rule engine is adopted, the confidence coefficient defaults to 1.0, and if the same abnormal behavior continuously appears for more than two times, additional deduction is triggered; and outputting the data quality scores, and performing visual marking or operation prompt on the original eye movement records according to the score grades, wherein the score grades are classified into high quality of 90 and above, available from 70 to 89 and unreliable from less than 70.
  2. 2. The method of claim 1, wherein the pre-set rules engine comprises the following decision rules: when the eye closing duration is more than or equal to 2 seconds and the head pitch angle change is less than 5 degrees, judging that the eye closing is interfered; when the triaxial angular velocity RMS is greater than 5 °/S and the change in the yaw angle is greater than 15 °, it is determined that the attention is dispersed; When the pitch angle change is more than 10 ° and the resting state ratio is less than 30%, it is determined that the head is swaying.
  3. 3. The method of claim 1, wherein the preset withholding rules comprise: Eye closure interference single button 10 minutes; The head shakes for 8 minutes; The attention is dispersed for 7 minutes by a single button; when the same abnormal behavior continuously occurs for more than two times, the extra button is 15 minutes; Actual deduction value = base deduction value x confidence.
  4. 4. The method of claim 1, wherein the machine learning classifier is a random forest model, a support vector machine, or a lightweight neural network, the input of which is a 28-dimensional fusion feature vector, the output of which is a probability distribution of four types of abnormal behavior, and the corresponding class is determined when the maximum probability is not less than 0.85.
  5. 5. The method of claim 1, wherein the dimension of the fusion feature vector is 28 dimensions, wherein the eye movement feature is 12 dimensions including eye closure duration, blink frequency, pupil visibility, left and right eye pupil position standard deviation, eye jump amplitude mean and standard deviation, and the head movement feature is 16 dimensions including three axis angular velocity RMS, three axis acceleration standard deviation, pitch/yaw/roll angle change rate mean and standard deviation, resting state duty ratio, head shake entropy.
  6. 6. The method of claim 1 wherein prior to extracting eye movement data, the eye movement data is median filtered with a filter window of 5 frames and the head movement data is second order Butterworth low pass filtered with a cut-off frequency of 10Hz.
  7. 7. The method of claim 1, wherein the data quality score is calculated as a sliding time window having a window length of 1.0 seconds and a step size of 0.5 seconds and displayed superimposed under the original eye movement waveform in the form of a color-coded curve in the inspection software interface, wherein the green color indicates a score of 90 or more, the yellow color indicates a score of 70 or less <90, and the red color indicates a score of <70.
  8. 8. The method of claim 1, wherein when the data quality score is an unreliable level, the system automatically pops up a prompt that "[ time period ] data quality is low (score = [ concrete value ]), suggesting a retest", and optionally pauses the current stimulation procedure.
  9. 9. The method of claim 1, wherein the method employs a rule engine path when deployed on an embedded device and a machine learning classifier path when deployed on a high performance workstation, both sharing the same feature extraction module and scoring output interface.
  10. 10. The method of claim 1, wherein the structured evaluation results generated by the method are derived in JSON or XML format, including time stamps, abnormal behavior categories, confidence levels, data quality scores, and clinical usage recommendations, and are invoked by the electronic medical record system.

Description

Abnormal behavior identification and data quality assessment method in vestibular function examination Technical Field The invention belongs to the technical field of crossing of medical instruments and artificial intelligence, and particularly relates to an abnormal behavior identification and data quality assessment method based on eye movement and head movement multi-mode data fusion in a vestibule function inspection process, which is suitable for data reliability judgment and clinical decision support in vestibule function detection scenes such as a cold and hot test, a video head pulse test (vHIT), dynamic vision target tracking and the like. Background Vestibular function examination is a core means for diagnosis of dizziness and related balance disorder diseases, and the accuracy of the vestibular function examination is highly dependent on active coordination of a tested person in the test process, namely, the vestibular function examination needs to keep eyes to keep looking at a target, the head is static, and random head turning or eye closure is avoided. However, in clinical practice, non-cooperative behavior often occurs due to discomfort, tension or distraction of the subject, resulting in distortion of eye movement records, which seriously affects the computational reliability of the eye shake parameters (e.g., slow phase velocity, gain, symmetry). Currently, the mainstream commercial devices already have basic interference monitoring capability, but there are significant limitations: The VN415 system of Interacoustics company (user manual rev.2022) detects eye closure events by infrared eye-tracker, marks red warning bars on the waveform when continuous eye closure exceeds 2 seconds, but does not interrupt the flow, nor does it evaluate whether the segment is available for diagnosis; The VisualEyes TM platform (product white paper, 2021) of Synapsis company integrates a head camera, can detect head offset of >10 degrees, but is only used for post playback prompt, and lacks real-time fusion analysis capability; The Micromedical system VisualEyes VNG supports IMU headset modules, but its exception handling is still based on independent thresholds (e.g., angular velocity >5 °/s) and does not model eye movement in combination with head movement characteristics, failing to identify compound interference behavior (e.g., slight head movement with line of sight shift). Academic research has also been explored, but the clinical landing problem is not solved: Zhang et al (IEEE Transactions on Biomedical Engineering, 2021) proposes to combine eye movement with IMU recognition driving distraction, but its model is designed for healthy drivers, does not take into account spontaneous eye-shake disturbances of vestibular disease patients, and does not output a data quality score; Lee & Kim (Journal of Vestibular Research, 2022) uses deep learning to reject the head too fast fragment in vHIT, but ignores eyelid status, and cannot handle pupil loss due to eye closure; Domestic patents CN1 14306287a (2022) and CN1 13974789a (2022) are used for cognitive assessment and parkinson's disease status monitoring, respectively, and are not applied to vestibular examination scenes, nor are eye movement-head movement fusion involved for data quality control. In order to improve the automation level, the applicant previously proposed an intelligent cooperative control system and method for vestibular function examination (chinese invention patent application number 202511870093.8). The system realizes multi-source signal time sequence coordination by synchronously controlling cold and hot stimulus, a visual target, a body position and a voice module, and can automatically pause and prompt the flow when the eye is detected to be closed for more than 2 seconds or the angular speed is detected to be more than 5 degrees/s, thereby effectively reducing the completely invalid test. However, as can be seen from the above, the prior art (including the applicant's existing patents) still faces three common bottlenecks: (1) Abnormal recognition lacks semantic understanding that physiological blinks and pathological eye closure cannot be distinguished by relying on single sensor isolation threshold values, and compound interferences such as head movement, sight line deviation and the like cannot be recognized; (2) The data quality is unquantifiable, namely, only an event mark or flow interruption is provided, the credibility level of the eye movement fragment is not given, and a doctor still needs to subjectively judge the availability of the data; (3) The clinical decision support is weak, the output stays at the binary level of warning/suspending, and a complete decision chain from signal to behavior to grading to suggestion is not established. Therefore, an intelligent evaluation method capable of fusing multi-modal information, realizing semantic classification of abnormal behaviors and outputting quality scores of structured