A control system and method for generating and adaptively acquiring eye reversal response parameters

By constructing a multi-input response model to separate the complex inputs of eye torsion and head movement, the problem of unreliable parameters in eye reversal testing is solved, and reliable parameter estimation and test efficiency are achieved under non-standardized conditions.

CN123113394APending Publication Date: 2026-11-17FIRST HOSPITAL OF SHANXI MEDICAL UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202611542388.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-06
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing eye-roll test, it is difficult to separate the transient changes in head movement and the complex motion input caused by changes in the direction of gravity, which makes it impossible to reliably estimate the eye-roll response parameters. Furthermore, traditional methods lack effective processing of dynamic inputs, resulting in unreliable test results.

Method used

By constructing a multi-input response model and using eye torsion data and head motion data, the transient changes in gravity direction and head angular motion are separated, a multi-input response model is established, the target parameters of eye reversal response are estimated, and when the parameters do not meet the identifiability conditions, the amount of information is increased by supplementing motion input to ensure the reliability and identifiability of the parameters.

Benefits of technology

It achieves physically interpretable separation and reliable output of eye reversal response parameters under non-standardized conditions, improves testing efficiency and result reliability, reduces dependence on standardized operations, supports orientation specificity and head-to-truncation decoupling, and improves the independence and accuracy of parameter estimation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113394A_ABST
    Figure CN123113394A_ABST
Patent Text Reader

Abstract

This invention discloses a control system and method for generating and adaptively acquiring eye-reversal response parameters. The system acquires data on eye torsion, head movement, and trunk movement; constructs a first motion input representing the change in the direction of gravity relative to the head coordinate system, a second motion input representing the transient change in head angular motion, and a third motion input representing the relative motion between the head and trunk; establishes a multi-input response model and estimates at least two target parameters for the eye-reversal response; evaluates the identifiable state for each target parameter; when the identifiable condition is not met, generates supplementary acquisition commands or motion control signals based on the information gap; updates the model using supplementary data; determines the parameter uncertainty or reliability based on the quality of eye torsion observation and motion input; and performs parameter-level gating, outputting only the target parameters that meet the identifiable and reliable conditions. This invention can improve the stability and reliability of eye-reversal response parameter estimation under non-strictly standardized tilting movements.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of biomedicine and computer signal processing technology, specifically to a control system and method for generating and adaptively acquiring eye reversal response parameters. Background Technology

[0002] Ocular counter-roll (OCR) is the torsional eye movement response of the eyes around the visual axis that occurs in the opposite direction to the head tilt when the head is tilted in a gravitational field. Video eye-tracking technology makes it possible to obtain the eye rotation angle from eye images; inertial measurement units, optical attitude tracking, and other technologies can simultaneously obtain head posture and motion information. Based on these technologies, existing video eye rotation tests typically involve tilting the head or head-to-torso as a whole while the subject is looking at a fixed target, measuring the change in eye rotation before and after tilting, and characterizing the eye rotation response as the eye rotation angle, the ratio of the eye rotation angle to the head tilt angle, or similar indicators.

[0003] Existing technologies also disclose schemes for estimating eye torsion using head posture and gravity vectors, calculating vestibular-ocular reflex indices using eye movement and head movement data, determining the validity of the test within a predetermined head movement rate range, and modifying subsequent vestibular motor stimuli based on the current test results. For example, CN106575162B discloses estimating eye torsion based on head posture and gravity vectors to improve eye tracking; US11317861B2 discloses performing vestibular-ocular reflex tests when the head movement rate meets a predetermined range and providing feedback based on the obtained indices; WO2013151773A1 / US9795335B2 discloses adaptively selecting subsequent motor stimuli based on current parameter estimation in vestibular psychophysical tests to improve test efficiency. Existing video eye-rolling schemes also involve eye torsion sensors, head posture sensors, and bedside tilting operations.

[0004] However, in the actual process of manually tilting, holding, and returning to center, transient motion inputs formed by changes in head posture relative to the direction of gravity, head angular velocity, and angular acceleration usually coexist. In some implementation scenarios, the motion of the head relative to the torso will also introduce additional inputs related to head-torso relative motion. Even if the final head tilt angle is the same, different motion trajectories, velocities, accelerations, dwell times, and pitch / yaw crosstalk will create different time-series inputs. If only a fixed time window is selected to calculate the eye torsion angle or simple gain, the information in the actual motion trajectory cannot be fully utilized, and the dynamic components may cause contamination with quasi-static quantization.

[0005] On the other hand, traditional quality control focuses more on observational conditions such as image clarity, frame loss, and whether head angle or velocity falls within a preset range. Even if these conditions are met, when different motion inputs are highly correlated, the amplitude of a certain type of input variation is insufficient, or a certain dynamic phase is not adequately covered, the multiple response parameters to be estimated may still be mathematically unstable and separable. Such "parameter indistinguishability" cannot be reliably detected through ordinary image quality scoring or motion range judgment.

[0006] Furthermore, existing adaptive vestibular tests typically modify subsequent motion stimuli based on subjective motion perception responses or overall test estimates, without addressing the objective response of ocular torsion by combining gravity-related inputs and transient angular motion inputs to establish a multi-input response model. Moreover, they do not predict the degree of improvement in parameter information content or covariance by different supplementary motion inputs based on the information gaps in the parameter subspace for specific ocular torsion response parameters that do not meet the identifiable conditions in the model, and then perform supplementary data acquisition.

[0007] Existing solutions often lack a mechanism to propagate uncertainties such as eye torsion observation errors, head motion measurement errors, time synchronization errors, and model residuals to the target parameters of each eye's inversion response. As a result, a situation may arise where a certain test can reliably obtain quasi-static gravity coupling parameters, but cannot reliably obtain dynamic coupling parameters or time constants. The system may still uniformly classify the entire test as "valid" or "invalid," and cannot provide reliable outputs based on individual parameters.

[0008] Therefore, a technical solution is needed that can construct multiple motion inputs with clear physical meaning based on the actual collected eye torsion and motion time series, use a multi-input model to evaluate the identifiability of each target parameter, supplement the corresponding motion excitation in a targeted manner when the specific parameter information is insufficient, and implement parameter-level gating output after propagating the multi-source measurement quality to the parameter credibility. Summary of the Invention

[0009] The technical problems to be solved by this invention are: during eye reversal testing, how to separate motion inputs representing different physical sources and estimate multiple target parameters of eye reversal response based on the actual collected eye torsion response and motion trajectory without relying on strictly fixed tilt angles, speeds, and holding times; how to determine whether the current data has sufficient independent information for each target parameter; how to determine supplementary motion inputs that can increase the estimated information based on the information gap corresponding to at least one target parameter when the information of that parameter is insufficient; and how to convert the quality of eye torsion observation and motion input into the uncertainty or credibility of each target parameter so that the system only outputs parameters that meet the preset identifiable and credible conditions.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] This invention provides an eye torsion response parameter generation and adaptive acquisition control system, including an eye torsion data acquisition module, a head motion data acquisition module, and a processing module connected to the eye torsion data acquisition module and the head motion data acquisition module. The processing module is configured to:

[0012] 1. Acquire time-corresponding eye torsion response data and head movement data;

[0013] 2. Construct at least a first motion input and a second motion input based on head motion data, wherein the first motion input represents the change in the direction of gravity relative to the head coordinate system, and the second motion input represents the transient change in the head angular motion;

[0014] 3. Establish a multi-input response model based on eye torsion response data, the first motion input, and the second motion input, and estimate at least two target parameters of eye torsion response using the multi-input response model;

[0015] 4. For each of the target parameters of the eye reversal response, based on the evaluation quantity related to the amount of input information or the stability of parameter estimation of the multi-input response model, determine the corresponding parameter identifiable state respectively;

[0016] 5. When at least one eye reversal response target parameter does not meet the preset identifiable condition, based on the information gap corresponding to the eye reversal response target parameter that does not meet the preset identifiable condition, determine the supplementary motion input to increase the estimation information of the corresponding parameter, and generate a supplementary acquisition command or motion control signal corresponding to the supplementary motion input.

[0017] 6. Update the multi-input response model after obtaining supplementary acquisition data corresponding to the supplementary motion input;

[0018] 7. Based on the observation quality of at least the ocular torsion response data and the quality of the first motion input and / or the second motion input, determine the parameter uncertainty or confidence level of the target parameters of each eye's torsion response;

[0019] 8. Perform parameter-level gating on each eye reversal response target parameter, and only output eye reversal response target parameters that simultaneously meet the preset identifiable conditions and preset reliable conditions. For eye reversal response target parameters that do not meet the conditions, output supplementary data or unreliable status information.

[0020] Preferably, the system further includes a trunk motion data acquisition module, and the processing module constructs a third motion input based on the head motion data and trunk motion data. The third motion input represents the posture change and / or angular motion of the head relative to the trunk, and the third motion input, together with the first and second motion inputs, enters the multi-input response model to estimate the response parameters related to the head-trunk relative motion or reduce the confounding effect of the motion on other parameters.

[0021] This invention also provides a method for generating and adaptively acquiring eye reversal response parameters, comprising:

[0022] Acquire time-corresponding eye torsion response data and head movement data;

[0023] Construct at least a first motion input and a second motion input based on head motion data;

[0024] A multi-input response model is established based on ocular torsion response data, a first motion input, and a second motion input, and at least two target parameters for ocular torsion response are estimated.

[0025] For each eye's reversal response target parameter, the parameter identifiability state is determined based on the evaluation quantity related to the input information or parameter estimation stability of the multi-input response model.

[0026] When at least one eye reversal response target parameter does not meet the preset identifiable condition, supplementary motion input is determined based on the corresponding information gap to increase the estimated information content of the eye reversal response target parameter, and supplementary acquisition command or motion control signal is generated.

[0027] Acquire supplementary ocular torsion response data and supplementary motion data corresponding to the supplementary motion input, and update the multi-input response model;

[0028] The parameter uncertainty or confidence level of the target parameters of each eye reversal response is determined based on at least the quality of the eye torsion observation and the quality of the first motion input and / or the second motion input.

[0029] Each eye reversal response target parameter is gated separately, and only eye reversal response target parameters that meet preset identifiable conditions and preset reliable conditions are output. For eye reversal response target parameters that do not meet the conditions, supplementary data or unreliable status information is output.

[0030] The core innovative logic of this invention can be summarized as follows: based on physically interpretable multi-input modeling, using parameter-level identifiability as the criterion, using information gap targeted supplementation as the means, and using multi-source uncertainty propagation and parameter-level gating as output guarantees, a closed loop of "acquisition-modeling-evaluation-supplementation-update-gated output" is formed.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. This invention does not simply improve the accuracy of a certain OCR indicator, but rather achieves physically interpretable separation of eye reversal response parameters, parameter-level identifiable diagnosis, information gap-driven directional adaptive supplementary sampling, and reliable output at the parameter granularity under non-standardized, multi-source noise, and dynamic hybrid conditions, thereby improving testing efficiency, result reliability, and clinical / research usability.

[0033] 2. The present invention constructs the change of gravity direction relative to the head coordinate system, the transient angular motion of the head, and the relative motion of the head and trunk as the first, second, and third motion inputs, respectively, so that the otolith gravity response, semicircular canal dynamic response, and neck / head-trunk relative motion response are no longer mixed into a single OCR angle or a single gain, thereby improving the physical interpretability and independent estimation ability of the target parameters of each eye inversion response.

[0034] 3. This invention utilizes actual collected head movement time series to establish a multi-input response model, enabling artificial tilt trajectories with non-strictly fixed angles, speeds, and holding times to still serve as effective input information, reducing reliance on standardized operations and fixed testing procedures, and more closely reflecting bedside or natural movement scenarios.

[0035] 4. This invention, through constraints such as the physiological range of the semicircular canal dynamic time constant, the gravitational-linear acceleration response relationship of the otolith organs, and the prior distribution of vestibular central compensatory parameters, suppresses parameter drift to non-physiological regions, making parameter estimation under limited data more stable and more in line with vestibular physiological laws.

[0036] 5. This invention evaluates the identifiability of each target parameter based on the weighted information matrix condition number, minimum eigenvalue, and relative standard error, and can detect situations where "the image is clear and the motion range is qualified, but the parameters cannot be stably separated due to collinear inputs or insufficient coverage," thus avoiding misjudging the entire test as valid or invalid based solely on observation quality or motion range.

[0037] 6. This invention reduces aimless repetitive testing by selectively supplementing information gaps: When a target parameter does not meet the identifiable condition, the information gap is located based on the direction of the low-information feature vector of the weighted information matrix, and the information gain of the candidate supplementary motion input is predicted. The action that can increase the information content of the corresponding parameter is selected, thereby achieving targeted excitation for the direction of missing information, instead of redoing all tests, improving the efficiency and security of data collection.

[0038] 7. This invention forms a closed-loop update, rather than a one-time judgment: supplementary collected data is used to update the multi-input response model and perform parameter-level identifiability evaluation again; if the conditions are still not met, supplementary actions can be selected until all target parameters pass the gating or the stopping condition is met. This closed loop makes the acquisition process and parameter estimation mutually driven, rather than just providing feedback on whether the action has reached the preset range;

[0039] 8. The multi-source uncertainty propagation of the present invention avoids outputting values ​​that are "superficially accurate but actually unreliable": it propagates the observation quality of eye torsion, motion input error, time synchronization error, model residuals, etc. to the variance or confidence interval of the target parameter, so that the weight of samples with low-quality eye torsion, attitude uncertainty or large synchronization error is automatically reduced, thereby providing a reliable basis for each parameter.

[0040] 9. This invention features parameter-level gated output, improving the utilization rate of single test results: Gating is performed separately for each target parameter, outputting only parameters that simultaneously meet the identifiable and reliable conditions; some parameters are allowed to be output, while others are marked as "requires supplementary data" or "unreliable." This changes the traditional approach of uniformly determining the entire test as valid / invalid, allowing reliable parameters in the same test to be released immediately, while parameters with insufficient information continue to be collected.

[0041] 10. This invention supports orientation specificity, binocular tolerance, and head-to-truncation decoupling: orientation specific parameters for left and right tilt can be estimated separately and evaluated and gated independently; modeling can be performed separately or jointly for the left and right eyes, and observation variance can be adjusted using binocular consistency to identify monocular loss of lock; in the dual IMU implementation, head-to-truncation relative motion is used as a third physical input, which can further reduce the confounding between gravity-related inputs and head-to-truncation relative motion, and provide an independent estimation channel for neck-related response parameters. Attached Figure Description

[0042] Figure 1 This is a block diagram of the overall system structure of the present invention;

[0043] Figure 2 A schematic diagram showing the layout of the binocular eye torsion data acquisition module, the head motion data acquisition module, and the trunk motion data acquisition module;

[0044] Figure 3 A schematic diagram showing the relationship between the world coordinate system, head coordinate system, torso coordinate system, and left and right eye coordinate systems;

[0045] Figure 4 Flowchart for time synchronization and coordinate calibration of eye torsion data and head / trunk motion data;

[0046] Figure 5The flowchart for constructing the first motion input, the second motion input, and the third motion input;

[0047] Figure 6 A flowchart for the eye reversal multi-input response model and target parameter estimation;

[0048] Figure 7 The flowchart for the target parameter-level identifiability evaluation illustrates one implementation method of parameter-level identifiability evaluation.

[0049] Figure 8 A flowchart for selecting supplementary motion inputs and generating supplementary acquisition commands or motion control signals based on information gaps;

[0050] Figure 9 This diagram illustrates the propagation of eye torsion observation error, motion input error, and time synchronization error to the uncertainty of the target parameters.

[0051] Figure 10 A flowchart for parameter-level gating with target parameters;

[0052] Figure 11 This is a schematic diagram of the third motion input structure for the overall tilt of the head-to-torso and the tilt of the head relative to the torso in a dual motion sensor implementation.

[0053] Figure 12 A schematic diagram illustrating the changes in uncertainty of the information matrix or target parameters before and after data acquisition;

[0054] Figure 13 This is a schematic diagram of the output data structure for the eye reversal response parameter vector, parameter uncertainty, and parameter state.

[0055] In the diagram: 100, Eye Twist Response Parameter Generation and Adaptive Acquisition Control System; 110, Eye Twist Data Acquisition Module; 120, Head Motion Data Acquisition Module; 130, Trunk Motion Data Acquisition Module; 140, Time Synchronization Module; 150, Processing Module; 160, Interaction / Motion Control Module; 170, Storage Module; 180, Motion Execution Unit;

[0056] 111. Left eye sensing unit; 112. Right eye sensing unit; 121. Head motion sensing unit; 131. Trunk motion sensing unit; 151. Calibration and coordinate transformation unit; 152. Motion input construction unit; 153. Multi-input response model unit; 154. Identifiability evaluation unit; 155. Information gap and supplementary acquisition decision unit; 156. Model update unit; 157. Uncertainty evaluation unit; 158. Parameter gating and output unit; 161. Display / voice / tactile prompting unit; 162. Motion platform interface; Detailed Implementation

[0057] The following is combined Figures 1-13 The present invention will be further described below. It should be understood that the following embodiments are for illustrative purposes only and not for limiting the scope of protection of the present invention. Those skilled in the art can make equivalent substitutions for sensor types, model forms, thresholds, action libraries, and data processing flows without departing from the technical concept of the present invention.

[0058] In this specification, "eye reversal response parameters" refers to technical parameters that characterize the eye reversal response, estimated from eye torsion response data and motion data;

[0059] "Information gap" refers to the insufficient amount of information in certain parameter directions in the parameter space of the current collected data in the multi-input response model. It is manifested as the presence of eigenvalues ​​below the threshold in the information matrix, excessively large prediction covariance of target parameters, or excessively high correlation coefficients between target parameters.

[0060] "Information gain" refers to the expected increase in the information matrix or the reduction in the variance of the target parameter that a candidate supplementary motion input can generate.

[0061] I. System Overall Structure

[0062] like Figures 1-2 As shown, the eye torsion response parameter generation and adaptive acquisition control system 100 includes an eye torsion data acquisition module 110, a head motion data acquisition module 120, a trunk motion data acquisition module 130, a time synchronization module 140, a processing module 150, an interaction / motion control module 160, a storage module 170, and a motion execution unit 180.

[0063] The eye torsion data acquisition module 110 includes a left eye sensing unit 111 and a right eye sensing unit 112; the head motion data acquisition module 120 includes a head motion sensing unit 121; the trunk motion data acquisition module 130 includes a trunk motion sensing unit 131; the internal logic of the processing module 150 is divided into a calibration and coordinate transformation unit 151, a motion input construction unit 152, a multi-input response model unit 153, a identifiability evaluation unit 154, an information gap and supplementary acquisition decision unit 155, a model update unit 156, an uncertainty evaluation unit 157, and a parameter gating and output unit 158. The above units can be implemented by a single processor or distributed across a head-mounted device, a mobile terminal, a workstation, or a server.

[0064] The interaction / motion control module 160 includes a display / voice / tactile prompt unit 161 and a motion platform interface 162. The storage module 170 is used to save the original time series data, calibration parameters, model parameters, candidate action library, threshold, historical state and output results. The motion execution unit 180 can be an electric tilt chair, turntable, robot actuator or other device that can change the subject's head / body posture according to the control trajectory.

[0065] II. Data Acquisition and Time Synchronization

[0066] like Figure 4 As shown, the eye torsion data acquisition module 110 and the head motion data acquisition module 120 operate at a sampling rate no less than the bandwidth required for eye torsion dynamic analysis. In the video eye tracking implementation, the eye camera may use, for example, 100–250 frames per second; the head motion sensor may use 100–1000 Hz. The above sampling rate is only an example and can be adjusted according to the device performance.

[0067] The time synchronization module 140 establishes the time correspondence between eye torsion data and motion data using hardware triggering, shared clock, timestamp synchronization, network precision clock, or software post-registration. If software post-registration is used, the time offset is estimated using known synchronization events, eye movement transients caused by head movements, or test start / end trigger signals. The time synchronization module 140 can also output the estimated time offset Δt and its uncertainty σΔt for subsequent uncertainty propagation.

[0068] For data with different sampling rates, the system resamples all data to a unified time axis. During the resampling process, anti-aliasing filtering is performed first, and the original timestamps are preserved to track synchronization errors.

[0069] III. Coordinate Calibration

[0070] like Figure 3 and Figure 4 As shown, the system defines at least a world coordinate system W and a head anatomical coordinate system H. One axis of the world coordinate system is aligned with the direction of gravity; the head anatomical coordinate system defines the horizontal, forward / backward, and vertical axes. If a torso sensor is used, a separate torso coordinate system T is defined.

[0071] If the sensor coordinate system of the head motion sensing unit 121 is S H Then through the rotation matrix R HS Transform the sensor coordinates to head anatomical coordinates. R HS Obtained through static upright calibration, known roll / yaw calibration actions, multi-attitude least squares fitting, external optical markers, or mechanical calibration frames.

[0072] In one embodiment, the subject first maintains a normal, upright posture for 2–10 seconds to determine the vertical constraint of the head using the direction of gravity; then, a small-amplitude pure yaw or lateral roll is performed to estimate the remaining rotational degrees of freedom using the principal rotation axis. R is obtained by minimizing the error between the rotation axis measured by the sensors and the expected head anatomical axis. HS The torso sensor uses a similar method to obtain R. TQ .

[0073] During testing, the system monitors the sensor's slippage relative to the head. If the visual geometry based on the eye socket, bridge of the nose, or fixed landmarks is inconsistent with the posture predicted by the motion sensor for an extended period, or if the calibration residual exceeds a threshold, the uncertainty evaluation unit 157 increases the corresponding motion input covariance. In severe cases, the relevant parameters are marked as unreliable and a recalibration prompt is given.

[0074] IV. Observation of eyeball torsion

[0075] Taking video eye tracking as an example, near-infrared illumination and a camera are used to acquire eye images. The torsion angle time series y is obtained through pupil / iris segmentation, polar coordinate unwrapping, global or local texture matching, deep learning feature tracking, or other known methods. L (k), y R (k).

[0076] The system simultaneously saves the eye torsion observation quality for each frame or time window. If blinking, severe occlusion, glare, defocusing, or local tracking loss occurs, the weight of that time period is reduced instead of simply deleting all tests. In cases of temporary loss of lock in one eye, data acquisition continues while the other eye still has valid observation capabilities, and the joint consistency check is re-performed after both eyes recover.

[0077] V. Motion Phase Recognition and Input Construction

[0078] like Figure 5 and Figure 11 As shown, the motion input construction unit 152 constructs the first motion input u according to the head movement timing. g Second motion input u ω The system simultaneously divides the data into several states based on the absolute value of the head roll angular velocity, angular acceleration, and attitude stability, including an initial stationary phase, a dynamic roll phase, a plateau phase, and a righting phase. This segmentation is used for model basis function selection, motion coverage evaluation, or observation weight adjustment, but it is not required to use a fixed threshold segmentation as a necessary condition for this invention.

[0079] Let the world coordinate system be W, the head anatomical coordinate system be H, and the gravity vector in the world coordinate system be g. W The rotation matrix of the head relative to the world coordinate system is R. WH (k), then the gravity vector in the head coordinate system can be represented as:

[0080] ;

[0081] By g H (k) Construct the first motion input u g (k). For example, when the unit vector of the horizontal axis of the head anatomical coordinate system is e y At that time, it can be ordered that:

[0082] ;

[0083] The roll angle φ of the head relative to the direction of gravity can also be used. g (k), sinφ g (k) The lateral gravity projection or other equivalent function is used as the first motion input.

[0084] Second motion input u ω (k) is determined by the head roll angular velocity ω r (k), angular acceleration α r (k) The filtering state of angular velocity or a combination thereof. For example:

[0085] ;

[0086] Or first, ω r (k) Input a filter with preset or estimated dynamic characteristics to obtain the state x characterizing the transient history effect of angular motion. ω (k), then with x ω (k) is at least one component of the second motion input.

[0087] In implementations with trunk motion data, such as Figure 11 As shown, let the torso coordinate system be T, and the head and torso posture quaternions be q, respectively. H( k) and q T (k), to determine the relative posture of the head with respect to the torso:

[0088] ;

[0089] According to q HT (k) Obtain the head-to-torso relative roll angle θ HT (k), relative angular velocity ω HT (k) etc., and construct the third motion input u n (k).

[0090] VI. Multi-input response model

[0091] like Figure 6 As shown, the multi-input response model is used to describe the quantitative relationship between the ocular torsional response and different physical inputs on a unified time axis. This model can employ linear or nonlinear regression, dynamic basis functions, state space, recursive least squares, Kalman filtering, Bayesian estimation, or other models capable of outputting target parameters and their estimated stability information.

[0092] In a directly implementable model, for the ocular torsion observation y of the e-th eye... e (k), establish the following model: ;

[0093] Among them, b e For the baseline term; K g,e K represents the gravity-related coupling parameter. ω,e K represents the transient coupling parameter for angular motion. n,e γ is the head-to-torso relative motion coupling parameter; z(k) is the pitch, yaw, horizontal / vertical eye position, pupillary change, non-gravitational linear acceleration, or other disturbance / correction variables; e F is the coefficient vector of the interference / correction variables; g and F n K is a predefined linear, nonlinear, or hysteresis basis function. If the torso motion module is not configured, K is omitted. n,e F n [u n (k)] item.

[0094] Angular motion dynamic state x ω (k) A first-order discrete model is adopted:

[0095] ;

[0096] Where τ ω Let a(τ) be the dynamic time constant to be estimated or preset. ω ) and b(τ ω The time constant is determined by the sampling interval and the selected discretization method. Alternatively, multiple filter basis functions with fixed time constants can be used to replace a single time constant to be estimated, thus maintaining the linear parametric form of the model.

[0097] The parameter vector p to be estimated e Includes K g,e K ω,e τ ω K n,e b e and γ e At least two of them. For binocular joint modeling, the left and right eyes can have independent parameters, or some parameters can be set with common parameters, sign constraints, amplitude ratio constraints, or soft regularization constraints.

[0098] For ease of description, K will be referred to as such in the following text without distinguishing between eye types. g,e Abbreviated as K g , will K ω,e Abbreviated as K ω , will K n,e Abbreviated as K n .

[0099] The aforementioned vestibular physiological priors can be introduced through parameter prior distributions, regularization terms, constraint equations, or training data priors. Specifically, the physiological range constraint for the semicircular canal dynamic time constant is 5-7 seconds; the otolith organ gravity-linear acceleration response constraint is based on the specific force equation, separating the gravity component of the world coordinate system gravity vector in the head coordinate system from the non-gravity linear acceleration component of the head, i.e.: let the gravity vector in the world coordinate system be g. W The rotation matrix of the head relative to the world coordinate system is R. WH (k), then the gravitational component in the head coordinate system is g H (k)=R WH (k)^T g W Non-linear acceleration of the head due to gravity a H (k) Through the force equation f H (k)=a H (k)-g H (k) is obtained, where f H (k) represents the accelerometer measurement, thus separating the gravity-related component from the linear acceleration component; the prior distribution of the vestibular central compensation parameters is a Gaussian prior or a physiological range constraint, such as the central compensation time constant τ. c Using a mean of τ c0 The standard deviation is σ c The Gaussian prior is used, or the constraint is within the range of [0.1, 10] seconds. As an example, the vestibular physiological model may adopt the Fernández-Goldberg semicircular canal model, the otolith organ gravity-linear acceleration response model, or an equivalent model, but the present invention is not limited to specific model names.

[0100] VII. Parameter Estimation and Observation Weights

[0101] In one weighted estimation implementation, the eye torsion data acquisition module 110 provides the eye torsion observation variance σ for each sampling point or time window. y ²(k) or equivalent mass value q eye (k). Quality values ​​are obtained based on the proportion of the effective eye region, image sharpness, local matching consistency, eyelid / eyelash occlusion, reflection, binocular consistency, or other known metrics. If the input is a quality value, it is converted to observation variance using a calibration function.

[0102] For example, the settings are:

[0103] ;

[0104] Where σ² min To minimize the observation variance, c q The calibration coefficient q0 is used to avoid an excessively small denominator. Parameters are determined using an artificial eye setup, repeated sampling, or calibration data.

[0105] For the parameter vector p, weighted least squares or weighted nonlinear least squares are used:

[0106] ;

[0107] Where: w(k) = 1 / max[σ² eff (k), σ² floor ];σ² eff (k) represents the effective observation variance after considering the observation error of eye torsion and the motion input error; λ is the regularization coefficient; and L is the regularization matrix. For state-space realization, equivalent parameter estimates and covariance are obtained by Kalman filtering, extended Kalman filtering, unscented Kalman filtering, maximum likelihood, or Bayesian posterior estimation.

[0108] 8. Parameter-level identifiability evaluation

[0109] like Figure 7 As shown, to distinguish between "observation quality is acceptable" and "target parameters can be stably estimated," this invention performs a parameter-level identifiability evaluation on the target parameters. This evaluation targets the sensitivity and input independence of each target parameter in the multi-input response model, rather than simply determining whether the head movement is within a preset range.

[0110] In one direct implementation scheme, suppose the model predicts the parameter p. j The sensitivity is:

[0111] ;

[0112] The sensitivity matrix S is composed of all valid sampling points, and a weighted information matrix is ​​constructed based on the observation weights W: ;

[0113] Where λI is a nonnegative small quantity used for numerical stability, and the different parameter columns of S are first normalized according to a preset parameter scale or column norm to obtain the scale-normalized information matrix J. N This is to avoid distortion in the comparison of condition numbers caused by different units.

[0114] The system calculates J or J N rank, condition number κ, minimum eigenvalue λ min The evaluation metric is selected from at least one of the following: determinant, objective parameter conditional covariance, Fisher information content, or equivalent parameter estimation stability index. In this embodiment, the evaluation metric specifically adopts the scale-normalized information matrix J. N The condition number κ(J_N) and the smallest eigenvalue λ min and the relative standard error (RSE) of the target parameter jThe upper limit of the condition number κ(J_N) is set to a calibration value in the range of 50-200, and the minimum eigenvalue λ is... min The lower limit is determined based on simulation and bench tests to be a calibration value within the range of 0.01-0.1, and the relative standard error (RSE) is... j The upper limit is set to a calibration value within the range of 0.20-0.35. When κ(J_N) does not exceed the upper limit, λ... min Not lower than the lower limit and RSE j If the parameter does not exceed the upper limit, it is determined that the parameter meets the identifiable condition.

[0115] For parameter p j The covariance can be approximated by the inverse of the information matrix:

[0116] ;

[0117] And calculate the relative standard error:

[0118] ;

[0119] Where σ² r For the weighted residual variance, p scale,j This is the scaling constant used to handle near-zero parameters.

[0120] In one exemplary decision rule, when the scale-normalized information matrix is ​​full rank and κ(J) N (RSE not exceeding the preset threshold and target parameters) j If the threshold is not exceeded and the corresponding sensitivity energy or conditional information reaches the threshold, p will be... j The thresholds are marked as "identifiable"; when only some conditions are met, they are marked as "boundary"; and when the conditions are clearly not met, they are marked as "unidentifiable". Each threshold can be calibrated through simulation, benchtop repeated measurements, or repeatable experiments on healthy subjects.

[0121] The preset identifiable conditions can be dynamically adjusted based on the subject's condition or prior clinical information. The subject's condition or prior clinical information includes at least one of the following: vestibular pathological pattern prior, disease stage, tolerance, age, and fixation ability. For example, for subjects with acute vestibular syndrome or low tolerance, the dependence on rapid lateral rolling movements can be reduced, while the weight of stable plateau segments or slow ramps can be increased; for subjects suspected of having central compensatory differences, different identifiable thresholds can be applied to the dynamic coupling parameters and time constants. The above adjustments are only used for data acquisition control and parameter output gating, and do not output disease diagnostic conclusions.

[0122] Furthermore, identifiability evaluations are performed on the left-side specific parameters and the right-side specific parameters respectively. When only the left-side tilt data is sufficient while the right-side tilt data is insufficient, the system marks the right-side specific parameters as requiring supplementary data and selects supplementary motion inputs that can increase the coverage of the right-side tilt phase.

[0123] For example, in an implementation without a defined protection range, the parameters of the sensitivity matrix are first normalized, and then the upper limit of κ is set to a calibration value within the range of 50 to 200, thus setting RSE... j The upper limit is set to a calibration value within the range of 0.20 to 0.35. The above values ​​are only used to illustrate the executable determination method and do not constitute a limitation on the present invention.

[0124] IX. Locating Information Gaps

[0125] When at least one target parameter fails to meet the preset identifiable condition, the system does not directly output "redo test," but instead determines the direction of the missing information for that parameter. The information gap is determined based on the direction of the low-information eigenvector of the weighted information matrix, specifically: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] N Perform eigenvalue decomposition J N =VΛV^T, select the eigenvectors corresponding to eigenvalues ​​less than the preset eigenvalue threshold, and analyze their loadings on each parameter dimension in the parameter set U that does not meet the identifiable condition. The parameter dimension with the larger loading is the direction of missing information. As a supplement, it can also be verified by combining the predicted covariance of the target parameters or the correlation coefficient between the target parameters.

[0126] In one implementation, the set of parameters that do not meet the identifiability criteria is denoted as U. The scale-normalized information matrix J... N Perform eigenvalue decomposition J N =VΛV^T, select the eigenvector corresponding to the small eigenvalue, analyze its loading on each parameter dimension in U, thereby determining which parameter combination information is mainly missing in the current data. Alternatively, the larger diagonal elements and higher correlation coefficients in Cov(p) can be used to locate the target parameters that are difficult to separate. Based on the determined information gap, the system further determines the physical input dimension corresponding to the gap. For example, when the low-information eigenvector is in u g When the load is large in the corresponding parameter dimension, it indicates insufficient gravity-related input excitation; when the low-information feature vector is in u ω When the load is large in the corresponding parameter dimension, it indicates that the transient input excitation of angular motion is insufficient.

[0127] 10. Targeted Supplementary Data Collection for Decision Making

[0128] like Figure 8As shown, the supplementary acquisition decision of the present invention is not based on the subject's subjective judgment of the direction of movement, nor is it based solely on whether the current head movement speed or angle is qualified. Instead, it targets the target parameter in the objective eye torsion multi-input response model that does not meet the identifiable condition, and predicts the improvement of the information content or covariance of the candidate supplementary motion input on that parameter.

[0129] The system stores a set of candidate supplementary motion inputs A in storage module 170. Candidate actions include, but are not limited to: adding a stable tilting platform segment on one side; adding a slower roll ramp; adding a roll motion segment with a large change in angular velocity; adding a direction reversal segment; adding motion in the opposite direction to the original acquisition; adding head-to-torso tilt or head-relative-torso tilt when torso motion data is available; and parameterized tilt trajectories generated by an automatic motion platform.

[0130] For candidate action a, the sensitivity matrix S during the re-collection period is predicted based on the current parameter estimation, action trajectory template, and sensor noise level. a and weight W a The incremental prediction information is obtained as follows:

[0131] ;

[0132] And calculate the prediction information matrix after supplementary sampling: J a =J+ΔJ(a);

[0133] For each candidate action, calculate the prediction covariance Cov of the target parameter set U. a Minimum eigenvalue, D-optimal criterion, or other information gain metrics. An example action scoring function is:

[0134] ;

[0135] Where α j The importance weight of the target parameter is represented by c(a), which represents the penalty for action duration, movement amplitude, subject tolerance, or device execution cost. The system selects the candidate action with the highest score(a), or the shortest / smallest action that reduces the variance of the target parameter to a preset threshold.

[0136] The specific trajectory parameter ranges for each candidate maneuver are as follows: Left stable tilting platform segment: roll angle 15-30°, angular velocity ≤5° / s, duration 5-10s; Right stable tilting platform segment: roll angle -15 to -30°, angular velocity ≤5° / s, duration 5-10s; Slow roll ramp: angular velocity 2-10° / s, amplitude 20-40°; Fast roll ramp: angular velocity 20-60° / s, amplitude 15-30°; Reverse direction roll: left then right, 15-25° each, angular velocity 20-40° / s; Head-to-thigh overall tilt: head and torso tilt synchronously, relative angle ≤5°; Head-to-thigh tilt: head rolls relative to torso 10-20°, torso remains stable; Automatic tilting platform trajectory: generated by platform parameterization, amplitude 10-30°, frequency 0.1-0.5Hz. The above parameter ranges can be adjusted according to subject tolerance and safety constraints.

[0137] In a simpler rule implementation, a parameter-action mapping table is established. For example: when the gravity coupling parameter K... g When information is insufficient, priority should be given to adding slow tilting with sufficient gravity projection changes and including a stable platform; when the dynamic coupling parameter K ω or time constant τ ω When information is insufficient, prioritize adding motion segments containing significant changes in angular velocity and reversal of direction; when K g With K n When there is a high correlation and the system has a torso sensor, priority should be given to adding actions where the head-torso relationship is different from the current data. Rule implementation and information gain implementation can be used separately or in combination.

[0138] The interactive / motion control module 160 can convert selected supplementary motion input into on-screen text, graphics, sound, light, or tactile cues for the operator or subject to execute; it can also generate platform trajectory parameters or control signals and send them directly to the motion execution unit 180. The specific angle, speed, and duration of the supplementary data acquisition are limited by the equipment capabilities and preset safety constraints.

[0139] XI. Supplementary Data Acquisition and Model Update

[0140] After supplementary data acquisition begins, the system continues to acquire eye torsion and motion data, using a unified time reference with the existing data. After quality checks, coordinate transformation, and motion input construction, the supplementary data is combined with the original data to form an updated dataset. Processing module 150 then re-estimates the model parameters or updates the parameters and covariance using a recursive update method.

[0141] For the weighted least squares model, the original information matrix J is added to the incremental information ΔJ obtained from the supplementary sampling, and then the solution is obtained again. For the recursive least squares or state-space model, the posterior parameters and covariance are updated progressively as new samples arrive. After each update, the parameter-level identifiability evaluation is re-executed. If the target parameters still do not meet the conditions, the system selects supplementary actions again until the stopping condition is met. The stopping condition may include: all target parameters pass the gating, the maximum number of supplementary samplings is reached, the maximum total test time is reached, the equipment or operator actively terminates the test, or it is determined that the remaining parameters are not identifiable under the current hardware configuration.

[0142] XII. Motion Input Quality and Uncertainty Propagation

[0143] like Figure 9 As shown, the quality of eye torsion observation and the quality of motion input jointly affect the reliability of the target parameters. Motion input quality includes at least one or more of the following: head pose estimation covariance, gyroscope noise or bias, gravity / linear acceleration separation error, sensor coordinate calibration residual, head-mounted device slip index, and time synchronization error.

[0144] In one implementation of first-order error propagation, if the local gradient of the model with respect to the motion input vector u(k) is G... u (k)=∂ŷ / ∂u, and the motion input covariance is Σ. u (k), the standard deviation of the camera-motion sensor time synchronization error is σ. t (Unit: seconds), the time derivative of the model prediction is dŷ / dt, then the effective observation variance can be constructed:

[0145] ;

[0146] Where σ² model Used to represent unmodeled error or residual noise. The above σ² eff (k) Enter the weighted estimation and parameter covariance calculation, automatically reducing the weight of samples with low-quality eye torsion, uncertain posture, or large synchronization error, and inputting the motion covariance Σ. u (k) Construct based on sensor noise parameters, calibration residuals, and attitude estimation covariance, for example, Σ u (k)=diag(σ 姿态 2 (k), σ 角速度 2 (k), σ 加速度 2 (k), σ 同步 2 (k)), where each component is determined by the gyroscope noise variance, accelerometer noise variance, attitude estimation covariance, and time synchronization error variance of the head motion sensing unit, respectively. The model's local gradient G to the motion input... u(k) = ∂ŷ / ∂u can be obtained through analytical differentiation or numerical difference. The model residual variance σ model 2 Sliding window variance estimation of model-predicted residuals.

[0147] In another embodiment, uncertainty is propagated via Monte Carlo: the original data is perturbed and sampled M times according to the eye torsion observation error distribution, head posture covariance, and time offset distribution; each time, the motion input is reconstructed and the parameters are estimated; the parameter variance, percentage position confidence interval, or posterior distribution is calculated based on the M sets of parameter samples. M is set to, for example, 100 to 2000 times depending on computational resources.

[0148] For state-space or Bayesian models, the filtered / smoothed posterior covariance, Hessian approximation, or sampled posterior can be directly used as the uncertainty of the target parameter. All of these different methods are used to establish the technical correlation between "raw measurement quality - model estimation - target parameter uncertainty".

[0149] The eye torsion data acquisition module can also output horizontal eye position, vertical eye position, and / or gaze quality scores. The processing module identifies saccades, smooth tracking, or optokinetic nystagmus remnants based on the horizontal eye position, vertical eye position, and / or gaze quality scores, and uses them as at least a portion of the covariate z(k) in the multi-input response model, and / or to adjust the eye torsion observation variance σ. y ²(k) or mass value q eye (k). This can reduce the interference of fixation instability, residual saccades, or OKN on the estimation of eye reversal response parameters.

[0150] 13. Parameter-level credibility gating

[0151] like Figure 10 As shown, the processing module 150 determines the identifiable state and the reliable state for each target parameter. Parameter-level gating considers at least simultaneously whether the parameter meets the preset identifiable condition and whether the parameter uncertainty or reliability meets the preset reliable condition. Preferably, factors such as the number of valid samples, action coverage, model residuals, outlier ratio, binocular consistency, and number of resamples may also be considered.

[0152] In one implementation, for parameter p j Define states: valid, requires additional data, and unreliable. If p j It meets the identifiability condition, and its confidence interval width, relative standard error, or posterior confidence level meets the preset conditions. The preset identifiability condition is: the condition number κ(J) of the weighted information matrix. N )≤200, minimum eigenvalue λ min ≥0.01, Relative Standard Error (RSE) j≤0.35; The preset confidence condition is: the width of the 95% confidence interval of the target parameter does not exceed 30% of the nominal value of the parameter, or the posterior confidence is not less than 0.8. At the same time, if the model residual diagnosis passes, the status is valid and the value and its uncertainty are output; if the main reason is insufficient current information and there is still executable supplementary motion input, the status is that supplementary data is needed; if the confidence condition is not met due to continuous low observation quality, sensor slippage, model mismatch, or reaching the maximum number of supplementary samplings, the status is unreliable.

[0153] Therefore, during the same test, some parameters may have reached a valid state and been released, while other parameters still require additional data and should be collected again; the system does not need to uniformly determine the entire test as passed or failed.

[0154] Parameter-level gating is performed separately for left-direction-specific parameters and right-direction-specific parameters. This allows for a situation where, in the same test, the left-direction gravity coupling parameter is already valid and output, while the right-direction gravity coupling parameter or dynamic time constant is still incomplete or unreliable. The system can further output directional coverage indicators and left-right asymmetry indicators.

[0155] XIV. Output Parameters

[0156] like Figure 13 As shown, the target parameters for the eye reversal response output by the system are selected according to the specific model, including but not limited to: gravity-related coupling parameter K. g ; Angular motion transient coupling parameter K ω Dynamic time constant τ ω Head-to-torso relative motion coupling parameter K n ; response parameters in the left / right tilt direction; parameters for left-right asymmetry; parameters for binocular conjugate; model residual indices; variance, standard error, confidence interval or posterior confidence for each parameter; and valid, data-needed, unreliable or equivalent states for each parameter.

[0157] The target parameters for eye reversal response output by the system also include at least one of the following: left-direction specific gravity-related coupling parameters, right-direction specific gravity-related coupling parameters, left-direction specific angular motion transient coupling parameters, right-direction specific angular motion transient coupling parameters, left-right asymmetry parameters, directional coverage index, fixation quality index, and binocular conjugacy index.

[0158] The output of this invention is the technical parameters and reliability status obtained from sensor data processing, and it does not require further disease diagnosis conclusions to be given in the system or method.

[0159] Example 1: Basic binoculars video + head motion sensor version

[0160] In this embodiment, the eye torsion data acquisition module 110 uses binocular near-infrared video cameras, and the head motion data acquisition module 120 uses an inertial measurement unit installed on a head-mounted device. The subject keeps their gaze fixed on a target in front, first standing upright for about 5 seconds, then tilting naturally to the left once and holding the tilt briefly, then returning to the center position and tilting naturally to the right once and holding the tilt briefly. The tilting amplitude is performed by the operator within a preset safety range and is not required to be exactly equal to a fixed angle.

[0161] The system obtained the time series of left and right eye torsions and head pose q. H , roll velocity ω r and corresponding quality information; construct u g and u ω K is estimated using a weighted dynamic model. g K ω and τ ω After initial parameter estimation, J was calculated. N ,κ,λ min and each parameter RSE. If K g Pass, and τ ω If the test fails, the system will not require the entire process to be repeated; instead, it will prompt the user to add a roll motion that includes a significant speed change and a return to center. After re-samples, the model will be re-estimated, and K will be output according to the parameters. g K ω τ ω And its trusted status.

[0162] Example 2: Dual motion sensor head-torso decoupled version

[0163] This embodiment adds a trunk motion sensing unit 131 to Embodiment 1, which is fixed to the sternum, upper back, or other positions that can represent trunk posture. The system calculates q. HT θ HT and ω HT And construct the third motion input u n .

[0164] First, a head-to-torso overall lateral tilt was measured, so that u g obvious changes and u n Relatively small; another head tilt relative to the torso was collected, so that u n A significant change has been made. The system no longer relies on the operator's textual labeling of the action type, but instead constructs the input based on the actual relative posture of the two motion sensors.

[0165] If the initial data contains u g with u n High collinearity, K g and K nThe conditional covariance is large. After identifying this information gap, the system selects actions from the candidate action library that can change the correlation between the two. K is then re-estimated after supplementary sampling. g K n Other target parameters.

[0166] Example 3: Information Gain-Driven Adaptive Acquisition

[0167] This embodiment uses a candidate action set A = {left stable platform, right stable platform, slow roll ramp, relatively fast roll ramp, direction-reversed roll, head-to-torso overall tilt, head-to-torso tilt}. For the parameter set U that does not currently meet the identifiable conditions, the input trajectory of each candidate action is simulated, and ΔJ(a) and the predicted Cov after re-sampling are calculated. a .

[0168] The system selects actions based on Score(a). If multiple actions can bring the target parameter to the threshold, the action with the shorter expected duration, smaller amplitude, or higher operator-preset priority is prioritized. This decision is based on the specific target parameters of the eye reversal multi-input model and does not rely on the subject's subjective perception of motion direction.

[0169] Example 4: Monte Carlo propagation of multi-source errors

[0170] This embodiment is used to calculate parameter uncertainty. The system stores σ for each eye torsion sample. y (k) Head posture covariance Σ q (k) and time synchronization error σ Δt The processor performs 500 Monte Carlo samplings: each time sampling eye torsional perturbations, attitude perturbations, and time shifts from the corresponding error distribution, reconstructing u g u ω And estimate the parameters.

[0171] Calculate the standard deviation of the parameters and the 2.5%–97.5% percentile range based on 500 parameter samples. If K… g The interval width satisfies the threshold while τ ω If not satisfied, then K g The flag is valid, τ ω The labeling requires additional data or is unreliable.

[0172] Example 5: Binocular error tolerance and temporary unilateral delocking

[0173] When the quality of the left eye deteriorates for several consecutive frames due to eyelid occlusion, the system increases the observation variance of the left eye or suspends updates for that eye, while continuing to use right eye and head motion data. After the left eye recovers, the joint residuals of both eyes are used to determine whether a baseline jump has occurred; if so, a new local baseline is established for the left eye or the uncertainty in that segment is increased. As long as the overall uncertainty of the final target parameter still meets the conditions, the system can retain the parameters that have passed.

[0174] Example 6: Automatic Tilt Platform Version

[0175] The subject is positioned in a controlled tilt chair or platform. The motion execution unit 180 receives the re-sampled trajectory generated by the interaction / motion control module 160. The system first executes a low-amplitude initial trajectory and estimates K online. g The dynamic parameters are then used to predict the information gain of multiple permissible trajectories based on the parameter information gaps. Finally, a trajectory that meets the equipment safety constraints and is expected to minimize the variance of the target parameters is selected for execution.

[0176] This implementation directly converts the supplementary sampling command into a physical execution control signal. However, the motion platform is not a necessary structure for the independent system solution of this invention, so the bedside manual execution method is not excluded.

[0177] Example 7: Offline Existing Data Processing Version

[0178] The system imports existing raw video eye-tracking and head-motion data, performing time alignment, coordinate calibration, motion input construction, multi-input model estimation, identifiability evaluation, and parameter gating. If offline data lacks sufficient parameter information, the system outputs "suggested supplementary motion input types" and the corresponding failed parameter status, without needing to perform real-time data acquisition. If the user subsequently imports new supplementary data, the model is updated and re-gated.

[0179] Example 8: Pathological Prior and Tolerance Adaptation

[0180] The system receives subject status or prior clinical information, such as vestibular pathological pattern priors, disease stage, and tolerability scores. Based on this information, the system filters rapid rolls or direction reversals from the candidate action library and adjusts the identifiability threshold and confidence criteria. Supplementary acquisition decisions are only used for acquisition control and parameter gating and do not output disease diagnosis conclusions.

[0181] Example 9: Fixation quality and eye position covariates

[0182] Fixation quality and eye position covariates. The system acquires horizontal eye position, vertical eye position, and fixation quality scores, identifies saccades, smooth tracking, or OKN residuals, and uses them as covariates or adjusts observation weights.

[0183] Example 10: Left-right directional coverage and asymmetric parameters

[0184] The system divides the data into left tilt stage and right tilt stage according to the direction of motion, estimates the left and right direction-specific parameters respectively, and independently evaluates the identifiability and reliability, outputting the direction asymmetry index.

[0185] Alternative implementation methods

[0186] The eye torsion data acquisition module 110 is not limited to video iris technology; the head motion data acquisition module 120 is not limited to an inertial measurement unit; and the multi-input response model is not limited to the aforementioned first-order dynamic model. As long as the system establishes an eye torsion response model based on motion inputs with different physical meanings, performs a stability evaluation of the model's target parameters based on identifiability or equivalent parameter estimation, selects supplementary motion inputs based on the information gaps of parameters that do not meet the conditions, and performs parameter-level gating after determining the reliability of the parameters based on the observation / input quality, other equivalent implementations can be adopted.

[0187] For example, the second motion input can be angular velocity, angular acceleration, state variables filtered by semicircular canal dynamics, frequency domain components, or predefined dynamic basis functions; discriminability can be achieved using rank, condition number, minimum eigenvalue, Fisher information, parameter covariance, Bootstrap stability, or Bayesian posterior width; the selection of the re-sampling action can be achieved using the D-optimal criterion, A-optimal criterion, E-optimal criterion, target parameter variance reduction, expected information gain, or rule base; uncertainty can be achieved using analytical propagation, Bootstrap, Monte Carlo, filtered covariance, or Bayesian posterior.

[0188] When using a neural state-space model, a combination of known neural network structures (such as RNN, LSTM, or Transformer) with the state-space model can be adopted. The training can be carried out by minimizing the time-series prediction error. The specific structure and training method can be determined through conventional experiments based on actual data.

Claims

1. An eye torsion response parameter generation and adaptive acquisition control system, comprising an eye torsion data acquisition module (110), a head motion data acquisition module (120), a trunk motion data acquisition module (130), and a processing module (150) connected to the eye torsion data acquisition module (110), the head motion data acquisition module (120), and the trunk motion data acquisition module (130), characterized in that, The processing module (150) is configured to: Acquire time-corresponding eye torsion response data, head movement data, and trunk movement data; At least a first motion input and a second motion input are constructed based on head motion data, and a third motion input is constructed based on head motion data and trunk motion data. The first motion input represents the change of the direction of gravity relative to the head coordinate system and is constructed based on the response relationship of the otolith organs to gravity and linear acceleration. The second motion input represents the transient change of head angular motion and is constructed based on the semicircular canal dynamics model. The third motion input represents the posture change and / or angular motion of the head relative to the trunk. A multi-input response model is established based on ocular torsion response data, a first motion input, a second motion input, and a third motion input. At least two target parameters for ocular torsion response are estimated by the multi-input response model. The multi-input response model adopts vestibular physiological prior constraints, which include at least one of the following: physiological range constraints of the semicircular canal dynamic time constant, constraints of the otolith organ gravity-linear acceleration response relationship, and prior distribution of vestibular central compensatory parameters. For each eye reversal response target parameter, based on the evaluation quantity related to the input information content or parameter estimation stability of the multi-input response model, the corresponding parameter identifiable state is determined respectively, wherein the evaluation quantity includes at least one of the following: the condition number of the weighted information matrix, the minimum eigenvalue of the weighted information matrix, and the relative standard error of the target parameter. When at least one eye reversal response target parameter does not meet the preset identifiable condition, supplementary motion input is determined to increase the information content of the corresponding parameter estimation based on the information gap corresponding to the eye reversal response target parameter that does not meet the preset identifiable condition, and supplementary acquisition command or motion control signal corresponding to the supplementary motion input is generated; wherein, the information gap is determined based on the direction of the low information feature vector of the weighted information matrix. The multi-input response model is updated after obtaining supplementary acquisition data corresponding to the supplementary motion input; Based on the observation quality of at least the ocular torsion response data and the quality of at least one of the first motion input, the second motion input, and the third motion input, determine the parameter uncertainty or confidence level of the target parameters of each ocular torsion response. Parameter-level gating is performed on each eye inversion response target parameter. Only eye inversion response target parameters that simultaneously meet the preset identifiable conditions and preset reliable conditions are output. For eye inversion response target parameters that do not meet the conditions, supplementary data or unreliable status information is output. It is allowed that some eye inversion response target parameters are output while others are not. The processing module (150) is further configured to perform at least one of the following: (a) Identify saccades, smooth tracking, or optokinetic residuals based on horizontal eye position, vertical eye position, and / or gaze quality data, and use at least one of horizontal eye position, vertical eye position, saccades, smooth tracking, or optokinetic residuals as a covariate of the multi-input response model and / or as observation weights for adjusting ocular torsion response data. (b) Based on head motion data, trunk motion data or directional information in motion input, the data is divided into left tilt stage and right tilt stage, the left direction-specific eye reversal response parameters and the right direction-specific eye reversal response parameters are estimated, and the identifiability evaluation and parameter-level gating are performed on the left direction-specific eye reversal response parameters and the right direction-specific eye reversal response parameters, respectively. (c) Obtain subject status or clinical prior information, and adjust the preset identifiable conditions, the preset reliable conditions, the candidate supplementary motion input library and / or the security constraints of the supplementary motion input accordingly, wherein the subject status or clinical prior information includes at least one of vestibular pathological pattern prior, disease stage, tolerance, age, and fixation ability, and does not output a disease diagnosis conclusion.

2. The system according to claim 1, characterized in that, The first motion input includes at least one of the following: gravity-related components transformed from the world coordinate system gravity vector to the head coordinate system via head posture transformation, roll angle of the head relative to gravity, and a sine function or equivalent function of the roll angle. Furthermore, the first motion input separates the gravity-related components from the non-gravity linear acceleration components of the head based on the otolith response relationship or compensates for the non-gravity linear acceleration. The second motion input includes at least one of the following: head roll angular velocity, head roll angular acceleration, and state variables processed by a semicircular canal dynamic model or a predefined dynamic filter. The semicircular canal dynamic model has a preset or physiologically constrained fluid dynamic time constant. The third motion input is constructed by performing relative operations on the head posture quaternion and the trunk posture quaternion to obtain the head-to-truncation relative roll angle and / or relative angular velocity.

3. The system according to claim 1, characterized in that, The multi-input response model includes a weighted regression model, a basis function model, a state-space model, a recursive least squares model, a Kalman filter model, or a Bayesian estimation model. The multi-input response model includes at least a gravity-related channel corresponding to the first motion input, an angular motion transient channel corresponding to the second motion input, and a head-to-thigh relative motion channel corresponding to the third motion input. The processing module uses weighted least squares, weighted nonlinear least squares, Kalman filtering, or Bayesian posterior estimation for parameter estimation and determines sample weights or observation covariance based on the observation quality of the ocular torsion response data. The vestibular physiological prior constraints are introduced through at least one of the following: parameter prior distribution, regularization term, constraint equation, or training data prior.

4. The system according to claim 1, characterized in that, The processing module predicts the information matrix increment of candidate supplementary motion inputs and calculates information gain index, minimum eigenvalue increment, D-optimal criterion index, E-optimal criterion index, or target parameter variance reduction based on the information matrix increment, selecting the candidate supplementary motion input that optimizes the objective function. The candidate supplementary motion inputs are selected from a candidate action library, which includes at least two of the following categories: left stable tilting platform segment, right stable tilting platform segment, slow roll ramp, fast roll ramp, direction reversal roll, head-to-thigh overall tilt, head relative to torso tilt, and automatic tilting platform trajectory. Furthermore, the processing module filters candidate supplementary motion inputs based on subject tolerance, prior clinical information, or safety constraints.

5. The system according to claim 1, characterized in that, The processing module propagates at least two of the following torsional observation variance, motion input covariance, time synchronization error, and model residual to the target parameter variance or confidence interval through first-order error propagation or Monte Carlo sampling: The parameter-level gating outputs a valid, data-required, or unreliable status for each target parameter. The processing module also models the left and right eyes separately or jointly, adjusts the observation variance using the consistency between the left and right eyes or identifies monocular loss of lock, and outputs a binocular conjugate index. The processing module also outputs parameter values, parameter uncertainties, identifiable states, and reliable states for the left-direction specific eye reversal response parameters and the right-direction specific eye reversal response parameters, and outputs left-right asymmetry indicators.

6. A method for generating and adaptively acquiring eye reversal response parameters, characterized in that, include: S1. Acquire time-corresponding eye torsion response data, head movement data, and trunk movement data; S2. Construct at least a first motion input and a second motion input based on head motion data, and construct a third motion input based on head motion data and trunk motion data, wherein the first motion input represents the change of the direction of gravity relative to the head coordinate system and is constructed based on the response relationship of the otolith organs to gravity and linear acceleration, the second motion input represents the transient change of head angular motion and is constructed based on the semicircular canal dynamics model, and the third motion input represents the posture change and / or angular motion of the head relative to the trunk. S3. Based on ocular torsion response data, a first motion input, a second motion input, and a third motion input, establish a multi-input response model and estimate at least two target parameters for ocular torsion response. The multi-input response model adopts vestibular physiological prior constraints, which include at least one of the following: physiological range constraints of the semicircular canal dynamic time constant, constraints of the otolith organ gravity-linear acceleration response relationship, and prior distribution of vestibular central compensatory parameters. S4. For each eye reversal response target parameter, determine the parameter identifiable state based on the evaluation quantity related to the input information content or parameter estimation stability of the multi-input response model, wherein the evaluation quantity includes at least one of the following: the rank of the sensitivity matrix or information matrix of the multi-input response model, the condition number of the weighted information matrix, the minimum eigenvalue of the weighted information matrix, the Fisher information content, the prediction covariance of the target parameter, the relative standard error, or the confidence interval width. S5. When at least one eye reversal response target parameter does not meet the preset identifiable condition, supplementary motion input is determined based on the corresponding information gap to increase the estimated information content of the eye reversal response target parameter, and supplementary acquisition command or motion control signal is generated; wherein, the information gap is determined based on the direction of the low information feature vector of the weighted information matrix; S6. Obtain supplementary ocular torsion response data and supplementary motion data corresponding to the supplementary motion input, and update the multi-input response model; S7. Determine the parameter uncertainty or confidence level of the target parameter of each eye reversal response based on at least the ocular torsion observation and the quality of at least one of the first motion input, the second motion input and the third motion input; S8. Gating is performed on each eye inversion response target parameter. Only eye inversion response target parameters that meet the preset identifiable conditions and preset reliable conditions are output. For eye inversion response target parameters that do not meet the conditions, supplementary data or unreliable status information is output. Some eye inversion response target parameters are allowed to be output while others are not. The method further includes at least one of the following: Based on horizontal eye position, vertical eye position, and / or fixation quality data, identify saccades, smooth tracking, or optokinetic nystagmus remnants and use them as covariates and / or adjust observation weights; Based on directional information, the left tilt stage and the right tilt stage are divided. The left-direction-specific eye reversal response parameters and the right-direction-specific eye reversal response parameters are estimated, and the identifiability evaluation and parameter-level gating are performed respectively. The system acquires the subject's status or prior clinical information and adjusts the pre-defined identifiable conditions, pre-defined reliable conditions, candidate supplementary motion input library, and / or supplementary motion input safety constraints accordingly, without outputting a disease diagnosis conclusion.

7. The method according to claim 6, characterized in that, The identifiable state of the parameters is determined by calculating the sensitivity matrix and the weighted information matrix, and based on at least one of rank, condition number, minimum eigenvalue, Fisher information, parameter covariance, or relative standard error; the information gain of candidate supplementary motion inputs is predicted, and the candidate supplementary motion input that maximizes the decrease in target parameter variance, maximizes the increase in minimum eigenvalue, or maximizes the D-optimal criterion index is selected; at least two of the eye torsion observation variance, motion input covariance, time synchronization error, and model residuals are propagated to the target parameter variance or confidence interval.

8. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, it causes the electronic device to perform the method of claim 6 or 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of claim 6 or 7.

Citation Information

Patent Citations

  • Enabling eye tracking based on dynamic eye torsion on computing devices

    CN106575162B

  • Vestibular-ocular reflex test and training system

    US11317861B2

  • Data collection for vestibulogram construction

    US9795335B2

  • Data collection for vestibulogram construction

    WO2013151773A1