Systems and methods for neural network-based eye movement data tagging and classification

Abstract

Methods, systems, and devices for feature extraction and assessment of visual testing are provided. One method comprises generating visual stimuli configured to elicit an eye movement data in a subject; receiving the eye movement data comprising optokinetic reflex (OKR) responses; selecting a classification model by performing a probabilistic analysis on a set of selection parameters applied to the eye movement data; segregating the plurality of eye movement data to generate tagged classifications by assigning predefined tags to measurable parameters of the eye movement data and converting the tagged classifications into normalized numerical values to generate classified data; generating feature data by extracting the tagged classifications from the classified data based on relevancy parameters associated with the classification model; determining a condition when a probability score meets a decision threshold; and generating administration data for the condition based on a pre-defined database configured with treatment protocols specific to the condition.

Description

Title: SYSTEMS AND METHODS FOR NEURAL NETWORK-BASED EYE MOVEMENT DATA TAGGING AND CLASSIFICATIONCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of United States Provisional Patent Application No. 63 / 749,348 filed Jan. 24, 2025, and the entire contents of United States Provisional Patent Application No. 63 / 749,348 are hereby incorporated herein in its entirety.FIELD

[0002] The embodiments described herein generally relate to systems and methods for analyzing eye movement data by artificial intelligence models, and in particular, to the extraction, tagging, and classification of eye movement features, including optokinetic response patterns, within a neural network framework.BACKGROUND

[0003] The following is not an admission that anything discussed below is part of the prior art or part of the common general knowledge of a person skilled in the art.

[0004] The optokinetic reflex (OKR) refers to an involuntary response of the human eye when it is presented with moving visual patterns. When exposed to stimuli like alternating striped lines moving in a particular direction, the eyes naturally follow the motion. The reflexive movements are often subtle, yet they can provide a wealth of information about underlying eye movement behavior. For example, certain patterns of involuntary tracking movements may be associated with changes in alertness, vestibular function, or even cognitive performance.

[0005] Many factors influence the recorded eye movements. Differences in the width, speed, or orientation of visual patterns can produce slight variations in the measured response. Environmental conditions, such as ambient lighting or the positioning of the subject’s head, may also alter the nature of the collected signals.

[0006] Although such differences may appear minor, they are frequently of interest in fields that require detailed analysis of eye movements. In some cases, small changes over time can indicate shifts in an individual’s response patterns. Understanding these shifts may be valuable for research applications, training programs, or long-term monitoring efforts.

[0007] Processing large volumes of eye-movement data poses significant technical challenges. Traditional methods often rely on manual inspections, fixed thresholds, or conventional signal processing techniques. These approaches may struggle when facing complex patterns or large datasets containing numerous subtle variations. Additionally, manual analysis is labor-intensive and can introduce subjectivity. Different observers may label the same movements inconsistently, leading to a lack of standardization. As the amount of data grows, human oversight becomes impractical, making it difficult to maintain consistent quality or scale to larger sample sizes.

[0008] Conventional signal processing methods can also be limited. Predefined filters or static algorithms might fail to capture all the nuances within the data. Small, irregular fluctuations that could hold valuable information may be overlooked. Such methods often cannot adapt effectively as conditions, stimuli, or individual responses change. Technical limitations can also arise because captured eye movement data is dependent on the stimulus protocol and acquisition validity. The responses recorded under different directions, speeds, colors / contrast, or ambient light (including darkness phases) may be non-comparable without explicit stimulus context. Additionally, real-time systems may waste compute and storage on unusable samples, or may lose diagnostically relevant features when relying only on raw frames or one-size-fits-all feature sets. These issues can lead to inconsistent tagging, unreliable scoring across stimulus blocks and sessions, and poor scalability for on-device or near-edge processing.

[0009] Accordingly, there is a need for alternative systems and methods that can address the inefficiencies of the conventional systems.SUMMARY

[0010] This summary is intended to introduce the reader to the more detailed description that follows and not to limit or define any claimed or as yet unclaimed invention. One or more inventions may reside in any combination or sub-combination of the elements or process steps disclosed in any part of this document including its claims and figures.

[0011] In one aspect, in accordance with the teachings herein, there is provided at least one embodiment of a device for feature extraction and assessment of a condition of a subject by performing visual testing on the subject, the device comprising: a processor configured to: generate a plurality of visual stimuli configured to elicit eye movement in asubject, the plurality of visual stimuli comprising one or more patterns; display the plurality of visual stimuli to the subject; receive image data capturing any eye movements and any eye lid movements of the subject due to presentation of the plurality of visual stimuli where the image data comprises optokinetic reflex (OKR) responses; process the image data to obtain eye movement data including at least one measurement of eye movement and / or at least one measurement of eye lid movement; select a trained classification model by performing a probabilistic analysis on a set of selection parameters applied to the eye movement data, the set of selection parameters comprising characteristics of a plurality of eye movement data; segregate the plurality of eye movement data to generate tagged classifications by assigning predefined tags to measurable parameters of the eye movement data and converting the tagged classifications into normalized numerical values to generate classified data; generate feature data by extracting the tagged classifications from the classified data based on relevancy parameters associated with the trained classification model, the relevancy parameters comprising condition-specific tags; and determine or assess the condition based on a probability score.

[0012] In at least one embodiment, the processor is configured to generate the probability score using by processing the tagged classifications in the feature data with assigned weights to identify patterns indicative of the condition and optionally comparing the probability score to a threshold.

[0013] In at least one embodiment, the processor is configured to generate administration data for the condition based on a pre-defined database configured with treatment protocols specific to the condition, the administration data comprising treatment recommendations.

[0014] In at least one embodiment, the processor is configured to provide the treatment recommendations based on the determined or assessed condition to a person who performed the visual testing.

[0015] In at least one embodiment, the plurality of visual stimuli is generated on a display screen and the image data is captured by an image sensing device, and wherein the one or more patterns may have one or more speeds, one or more directions, and one or more colors.

[0016] In at least one embodiment, the eye movement data comprises any combination of amplitude, frequency, velocity, pursuit gain, interruptions, symmetry, and directionality of eye movements.

[0017] In at least one embodiment, the selection parameters provide an indication of a potential condition, and the selection parameters include any combination of saccadic movements, smooth pursuit, reflexive gaze stabilization, pursuit gain, and interruption frequency, and wherein the selection parameters are determined by user input.

[0018] In at least one embodiment, the predefined tags comprise any combination of nystagmus, saccades, smooth pursuit, look-to-stare ratios, and palpebral fissure area (PFA) measurements.

[0019] In at least one embodiment, the relevancy parameters are identified by performing a probabilistic analysis on training data comprising labeled datasets, and wherein the probabilistic analysis involves correlating the relevancy parameters with the condition.

[0020] In at least one embodiment, the processor is configured to generate inferential data identifying a plurality of additional condition-specific tags based on the analysis of node weights and correlation patterns by the neural network, the inferential data providing probabilistic relevance of untagged parameters to the condition.

[0021] In at least one embodiment, the condition comprises a neurological condition or a physiological condition wherein the neurological condition includes dementia, Parkinson’s disease, vestibular disorders, cranial nerve palsies and stroke, and wherein the physiological condition includes intoxication and fatigue.

[0022] In another aspect, in accordance with the teachings herein, there is provided at least one embodiment a method for feature extraction and assessment of a condition of a subject by performing visual testing on the subject, the method comprising: generating a plurality of visual stimuli configured to elicit eye movement in a subject, the plurality of visual stimuli comprising one or more patterns; displaying the plurality of visual stimuli to a subject via a display; receiving image data capturing any eye movements and any eye lid movements of the subject due to presentation of the visual stimuli where the image data comprises optokinetic reflex (OKR) responses; processing the image data to obtain eye movement data including at least one measurement of eye movement and / or at leastone measurement of eye lid movement; selecting a trained classification model by performing a probabilistic analysis on a set of selection parameters applied to the eye movement data, the set of selection parameters comprising characteristics of a plurality of eye movement data; segregating the plurality of eye movement data to generate tagged classifications by assigning predefined tags to measurable parameters of the eye movement data and converting the tagged classifications into normalized numerical values to generate classified data; generating feature data by extracting the tagged classifications from the classified data based on relevancy parameters associated with the trained classification model, the relevancy parameters comprising condition-specific tags; and determining or assessing the condition based on a probability score.

[0023] In at least one embodiment, the method comprises generating the probability score by processing the tagged classifications in the feature data with assigned weights to identify patterns indicative of the condition and optionally comparing the probability score to a threshold.

[0024] In at least one embodiment, the method further comprises generating administration data for the condition based on a pre-defined database configured with treatment protocols specific to the condition, the administration data comprising treatment recommendations.

[0025] In at least one embodiment, the method includes providing the treatment recommendations based on the determined or assessed condition to a person who performed the visual testing.

[0026] In at least one embodiment, the plurality of visual stimuli is generated on a display screen of an electronic device including a monitor, a visor or a VR headset and the plurality of image data is captured by an image sensing device, and wherein the one or more patterns may have one or more speeds, one or more directions, and one or more colors.

[0027] In at least one embodiment, the eye movement data comprises any combination of amplitude, frequency, velocity, pursuit gain, interruptions, symmetry, and directionality of eye movements.

[0028] In at least one embodiment, the selection parameters provide an indication of a potential condition, and the selection parameters include any combination of saccadicmovements, smooth pursuit, reflexive gaze stabilization, pursuit gain, interruption frequency.

[0029] In at least one embodiment, the predefined tags comprise any combination of nystagmus, saccades, smooth pursuit, look-to-stare ratios, and palpebral fissure area (PFA) measurements.

[0030] In at least one embodiment, the relevancy parameters are identified by performing a probabilistic analysis on training data comprising labeled datasets, and wherein the probabilistic analysis involves correlating the relevancy parameters with the condition.

[0031] In at least one embodiment, the method further comprises generating inferential data identifying a plurality of additional condition-specific tags based on the analysis of node weights and correlation patterns by the neural network, the inferential data providing probabilistic relevance of untagged parameters to the condition.

[0032] In at least one embodiment, the condition comprises a neurological condition or a physiological condition wherein the neurological condition includes dementia, vestibular disorders, cranial nerve palsies and stroke, and wherein the physiological condition includes intoxication and fatigue.

[0033] In another aspect, in accordance with the teachings herein, there is provided at least one embodiment of a device for detection and / or assessment of a condition of a subject via visual testing, the device comprising: an image sensor; a display; memory that stores a trained Artificial Intelligence (Al) model; and a processor that is communicatively coupled to the image sensor, the display and the memory, wherein the processor is configured to: generate a plurality of visual stimuli configured to elicit optokinetic reflex (OKR) responses in the subject; display the plurality of visual stimuli to the subject; receive image data capturing any eye movements and any eye lid movements of the subject due to presentation of the plurality of visual stimuli; process the image data to obtain eye movement data including at least one measurement of eye movement and / or at least one measurement of eye lid movement; process the image data using the trained Al model to generate a detected condition assessment and / or a condition assessment of the detected condition; and provide an indication of the detected condition assessment and / or the condition assessment of the detected condition.

[0034] In at least one embodiment, the measurements include a look-to-stare ratio, and / or palpebral fissure area (PFA) measurements.

[0035] In at least one embodiment, the Al model is one of a plurality of trained classification models that are associated with and trained for detecting and / or assessing a unique condition and one of the trained classification models is used as the Al model.

[0036] In at least one embodiment, the processor is configured to use trained selection model to perform a probabilistic assessment of the eye movement data to select one of the trained classification models to process the image data.

[0037] In at least one embodiment, the processor is configured to provide treatment recommendations based on the detected condition assessment and / or the longitudinal tracking of the detected condition.

[0038] In at least one embodiment, the condition comprises a neurological condition or a physiological condition wherein the neurological condition includes dementia, vestibular disorders, cranial nerve palsies and stroke, and wherein the physiological condition includes intoxication and fatigue.

[0039] In at least one of the device embodiments described herein, the processor is configured to perform validity gating for suppressing logging, streaming, and / or persistence of eye movement data unless predetermined validity criteria are satisfied including a valid gaze indication and valid bilateral pupil measurements.

[0040] In at least one of the method embodiments described herein, the method further comprises performing validity gating for suppressing logging, streaming, and / or persistence of eye movement data unless predetermined validity criteria are satisfied including a valid gaze indication and valid bilateral pupil measurements.

[0041] In another aspect, in accordance with teachings herein, there is provided at least one embodiment of a method for detection and / or assessment of a condition of a subject via visual testing. The method comprises generating a plurality of visual stimuli configured to elicit optokinetic reflex (OKR) responses in the subject; displaying the plurality of visual stimuli to the subject; receiving image data capturing any eye movements and any eye lid movements of the subject due to presentation of the plurality of visual stimuli; processing the image data to obtain eye movement data including at least one measurement of eye movement and / or at least one measurement of eye lid movement;processing the image data using the trained Al model to generate a detected condition assessment and / or a condition assessment of the detected condition; and providing an indication of the detected condition assessment and / or the condition assessment of the detected condition.

[0042] In another aspect, in accordance with the teachings herein, there is provided at least one embodiment of a non-transitory computer-readable storage medium storing program instructions that, when executed by a processor, cause the processor to perform a method defined according to any one of the embodiments described herein.

[0043] It will be appreciated that the foregoing summary sets out representative aspects of embodiments to assist skilled readers in understanding the following detailed description. Other features and advantages of the present application will become apparent from the following detailed description taken together with the accompanying drawings. It should be understood, however, that the detailed description and the specific examples, while indicating preferred embodiments of the application, are given byway of illustration only, since various changes and modifications within the spirit and scope of the application will become apparent to those skilled in the art from this detailed description.BRIEF DESCRIPTION OF THE DRAWINGS

[0044] For a better understanding of the various embodiments described herein, and to show more clearly how these various embodiments may be carried into effect, reference will be made, by way of example, to the accompanying drawings which show at least one example embodiment, and which are now described. The drawings are not intended to limit the scope of the teachings described herein.

[0045] FIG. 1 is a block diagram of a pattern assessment and administration system 100, according to an example embodiment.

[0046] FIG. 2 is a block diagram of a classification model selection system 200 200, according to an example embodiment.

[0047] FIG. 3 is a block diagram 300 of a classification model 300, according to an example embodiment.

[0048] FIG. 4 is a flow diagram 400 of a feature extraction and assessment method, according to an example embodiment.

[0049] FIG. 5 is a block diagram 500 of a user apparatus 500 for optokinetic stimulation, according to an example embodiment.

[0050] FIG. 6 is a block diagram 600 of a user apparatus 600 for optokinetic stimulation, according to an example embodiment.

[0051] FIG. 7 is a functional diagram 700 illustrating an example optokinetic stimulation and eye movement data, according to an example embodiment.

[0052] FIG. 8 is a functional diagram 800 illustrating an example optokinetic stimulation and eye movement data, according to an example embodiment.

[0053] FIG. 9 is a functional diagram 900 illustrating an example optokinetic stimulation and eye movement data, according to an example embodiment.

[0054] FIG. 10 is a functional diagram 1000 illustrating an example optokinetic stimulation and eye movement data, according to an example embodiment.

[0055] The skilled person in the art will understand that the drawings, described below, are for illustration purposes only. The drawings are not intended to limit the scope of the applicants' teachings in any way. Also, it will be appreciated that for simplicity and clarity of illustration, elements shown in the figures have not necessarily been drawn to scale. For example, the dimensions of some of the elements may be exaggerated relative to other elements for clarity. Further, where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or analogous elements.DESCRIPTION OF VARIOUS EMBODIMENTS

[0056] The headings and Abstract of the Disclosure provided herein are for convenience only and do not interpret the scope or meaning of the embodiments.

[0057] Various embodiments in accordance with the teachings herein will be described below to provide examples of at least one embodiment of the claimed subject matter. No embodiment described herein limits any claimed subject matter. The claimed subject matter is not limited to devices, systems or methods having all of the features of any one of the devices, systems or methods described below or to features common to multiple or all of the devices, systems or methods described herein. It is possible that there may be a device, system or method described herein that is not an embodiment of any claimed subject matter. Any subject matter that is described herein that is not claimed in this document may be the subject matter of another protective instrument, for example, acontinuing patent application, and the applicants, inventors or owners do not intend to abandon, disclaim or dedicate to the public any such subject matter by its disclosure in this document.

[0058] It will be appreciated that numerous specific details are set forth in order to provide a thorough understanding of the exemplary embodiments described herein. However, it will be understood by those of ordinary skill in the art that the embodiments described herein may be practiced without these specific details. In other instances, well-known methods, procedures and components have not been described in detail so as not to obscure the embodiments described herein. Furthermore, this description is not to be considered as limiting the scope of the embodiments described herein in any way, but rather as merely describing the implementation of the various embodiments described herein.

[0059] It should also be noted that the terms “coupled” or “coupling” as used herein can have several different meanings depending on the context in which these terms are used. For example, the terms coupled or coupling can have a mechanical, electrical or communicative connotation. For example, as used herein, the terms coupled or coupling can indicate that two elements or devices can be directly connected to one another or connected to one another through one or more intermediate elements or devices via an electrical element, an electrical signal, a light signal or a mechanical element depending on the particular context.

[0060] It should also be noted that, as used herein, the wording “and / or” is intended to represent an inclusive-or. That is, “X and / or Y” is intended to mean X or Y or both X and Y, for example. As a further example, the phrases “X, Y, and / or Z”, “any operable combination of X, Y and Z”, “X, Y, Z or any combination thereof” or “any combination of X, Y and Z” is intended to mean X, Y, Z, X and Y, X and Z, Y and Z, or X, Y and Z.

[0061] It should be noted that terms of degree such as “substantially”, “about” and “approximately” when used herein mean a reasonable amount of deviation of the modified term such that the end result is not significantly changed. These terms of degree should be construed as including a deviation of the modified term, such as by 1%, 2%, 5%, 10% or 15%, for example, if this deviation would not negate the meaning of the term it modifies.

[0062] The terms "including," "comprising" and variations thereof mean "including but not limited to," unless expressly specified otherwise. A listing of items does not imply that any or all of the items are mutually exclusive, unless expressly specified otherwise. The terms "a," "an" and "the" mean "one or more," unless expressly specified otherwise.

[0063] The terms "an embodiment," "embodiment," "embodiments," "the embodiment," "the embodiments," "one or more embodiments," "some embodiments," and "one embodiment" mean "one or more (but not all) embodiments of the present invention(s)," unless expressly specified otherwise.

[0064] Furthermore, the recitation of numerical ranges by endpoints herein includes all numbers and fractions subsumed within that range (e.g., 1 to 5 includes 1, 1.5, 2, 2.75, 3, 3.90, 4, and 5). It is also to be understood that all numbers and fractions thereof are presumed to be modified by the term “about” which means a variation of up to a certain amount of the number to which reference is being made if the end result is not significantly changed, such as 1%, 2%, 5%, 10% or 15%, for example.

[0065] The embodiments of the systems and methods described herein may be implemented in hardware or software, or a combination of both. These embodiments may be implemented in computer programs executing on programmable computers, each computer including at least one processor, a data storage system (including volatile memory or non-volatile memory or other data storage elements or a combination thereof), and at least one communication interface. For example, and without limitation, the programmable computers may be a server, network appliance, embedded device, computer expansion module, a personal computer, laptop, personal data assistant, cellular telephone, smart-phone device, tablet computer, a wireless device or any other computing device capable of being configured to carry out the methods described herein.

[0066] Program code may be applied to input data to perform the functions described herein and to generate output information. The output information is applied to one or more output devices, in known fashion.

[0067] Each program may be implemented in a high-level procedural or object oriented programming and / or scripting language, or both, to communicate with a computer system. However, the programs may be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language. Each suchcomputer program may be stored on a storage media or a device (e.g. ROM, magnetic disk, optical disc) readable by a general or special purpose programmable computer, for configuring and operating the computer when the storage media or device is read by the computer to perform the procedures described herein. Embodiments of the system may also be considered to be implemented as a non-transitory computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.

[0068] Furthermore, the system, processes and methods of the described embodiments are capable of being distributed in a computer program product comprising a computer readable medium that bears computer usable instructions for one or more processors. The medium may be provided in various forms, including one or more diskettes, compact disks, tapes, chips, wireline transmissions, satellite transmissions, internet transmission or downloads, magnetic and electronic storage media, digital and analog signals, and the like. The computer useable instructions may also be in various forms, including compiled and non-compiled code.

[0069] Optokinetic reflex (OKR) is an involuntary eye movement that occurs in response to moving visual stimuli. OKR helps stabilize the visual field during motion. Variations of OKR include horizontal and vertical OKR, which are triggered by different stimuli and produce distinct reflexive responses. Optokinetic nystagmus (OKN) comprises repetitive eye movements, alternating between smooth pursuit and rapid saccades. Optokinetic after-nystagmus (OKAN) is another form of a reflex, characterized by sustained eye movements following the end of a visual stimulus, often observed in darkness. The reflexes create measurable patterns of eye movement, which can show subtle variations caused by various factors and underlying conditions.

[0070] The term condition used herein for which the Al models are trained to detect may be a certain neurological condition or a physiological condition that is detectable through OKR testing. Examples of these neurological conditions include, but are not limited to, dementia, vestibular disorders, cranial nerve palsies and stroke. Examples of physiological conditions include, but are not limited to, intoxication and fatigue (e.g., sleepiness).

[0071] Eye movements are varied and include types such as saccades, smooth pursuit, and gaze-holding reflexes. Saccades are quick, discrete movements that redirect the gaze to a new point. Smooth pursuit movements allow the eyes to follow a moving object steadily. Reflexive eye movements, like those triggered by OKR, occur automatically to stabilize the gaze. The eye movement(s) generates distinct data patterns when recorded in response to controlled stimuli. The data patterns are further influenced by the characteristics of the stimulus, such as its speed, direction, and visual contrast.

[0072] Analyzing eye movement data produces large and complex datasets. The eye movement data may include measurements that have been made on images of movements of the subject’s eyes and / or eyelids due to visual testing. In some cases, the eye movement data may include this image data or a pre-processed version of it as well. The datasets include multiple parameters, such as the amplitude, frequency, velocity, and directional symmetry of movements. Small changes in these parameters can indicate deviations from expected patterns. It has been found by the inventors that changes in eye movement patterns can be used to understand underlying conditions that were not previously realized by those skilled in this field. The changes can be identified and analyzed using computational techniques. However, identifying such variations is not trivial as it involves separating meaningful signals from noise in the dataset, detecting subtle changes in patterns and categorizing them with high accuracy.

[0073] While recent advances in computational techniques, including machine learning and artificial intelligence, may be used to address these challenges, the inventors have determined that relevant features from signals obtained from a subject during Optokinetic testing may be used to improve accuracy and performance. Further by using certain processing techniques and computational models to extract relevant features, the inventors have determined it may become possible to identify subtle patterns that would remain hidden using traditional approaches and also to reduce / remove any subjectivity that may otherwise be present in analyzing these patterns.

[0074] However, the complexity and variability of eye movement data is challenging, which the inventors have determined, may be addressed by using a processing architecture that incorporates tagging and classification as described herein. Tagging and classifying eye movement data creates a structured framework for analyzing complexpatterns. By assigning tags to specific features, such as, but not limited to the amplitude of a saccade or the frequency of nystagmus, for example, the data can be organized for further processing. The tags can then be used to train machine learning models, enabling the recognition of similar patterns in new datasets.

[0075] Accordingly, in at least one embodiment, the present disclosure provides an artificial intelligence-based system for analyzing eye movement data to identify patterns. The system includes neural networks trained to process large volumes of eye movement data. The modular architecture includes components for pre-processing, filtering, and analyzing raw data collected during optokinetic reflex (OKR) tests. The architecture includes stages for data filtration, feature extraction, and model selection directed to identifying relevant patterns within the input data.

[0076] In at least one embodiment, the system may be implemented to provide a multimodel architecture configured to dynamically select the most suitable classification model, also referred to as an inference or Al model, for analyzing eye movement data for identifying a particular condition. The model selection may be based on a combination of selection parameters, including type of eye movement data, tagged features in the training dataset, and the characteristics of the optokinetic stimulus. The architecture may include a decision-making framework that evaluates the context of the input eye movement data and assigns the input data to the appropriate classification model. The multi-model architecture improves the system’s accuracy and efficiency in detecting subtle variations in eye movement pattern.

[0077] The classification model comprises a neural network architecture configured to process eye movement data. The classification model may include interconnected nodes organized into a plurality of data processing layers. The plurality of data processing layers includes an input layer, an inference layer, and an output layer.

[0078] The input layer of a classification model is configured to receive pre-processed eye movement data captured during optokinetic reflex (OKR) tests and generate tagged data, also referred to as classified data, comprising normalized and tagged eye movement data. The input layer receives pre-processed eye movement data. The pre-processed eye movement data includes any combination of speed, accuracy, reset delay,pursuit gain, interruptions, amplitude, frequency, symmetry, and directionality of eye movements, as well as auxiliary features such as gaze position and eyelid dynamics.

[0079] The accuracy may refer to the precision with which the subject’s eye movements align with the visual stimulus presented during testing. Accuracy is measured by tracking how closely the eyes follow the target without deviations. Tagged data capturing accuracy reflects the deviations as numerical values or labeled events indicating off-target tracking. The frequent off-target movements can be tagged as markers of neurological dysfunction such as, but not limited to, dementia or impaired motor control, for example.

[0080] The reset delay may refer to the time interval, measured in milliseconds, needed for the eye to return to its baseline or original position after completing a tracking movement, such as a smooth pursuit or a saccade. The tagged data for reset delay is generated by capturing and labeling the time lag between the completion of a movement and the initiation of the next movement. For example, a prolonged reset delay can be indicative of brainstem dysfunction, fatigue, or concussion.

[0081] The pursuit gain may refer to the ratio of the eye’s movement velocity to the velocity of the stimulus being tracked. Tagged data associated with pursuit gain is calculated during smooth pursuit movements and labeled to reflect the degree of alignment between eye velocity and stimulus velocity. Reduced pursuit gain, tagged as a marker, can be indicative of conditions such as, but not limited to, for example, dementia, Parkinson’s disease, or vestibular dysfunctions. For example, a pursuit gain below the standard threshold can indicate impaired smooth pursuit mechanisms.

[0082] The interruptions may refer to breaks or pauses in smooth pursuit movements, where the eye temporarily ceases tracking the stimulus and then resumes. Tagged data for interruptions captures the breaks or pauses as discrete instances, measuring their frequency and duration. Frequent or prolonged interruptions can be indicators of cognitive decline, fatigue, intoxication, or neurological impairments. For example, a subject with dementia may exhibit frequent interruptions in tracking a moving target.

[0083] The symmetry of eye movements may refer to the consistency in movement patterns across both eyes in response to a stimulus. The input layer tags symmetry by evaluating whether both eyes move with equal amplitude and directionality during tracking tasks. Tagged data for asymmetry captures deviations in synchronization between theeyes. For example, in cases of unilateral oculomotor palsy, one eye may lag behind the other. Asymmetrical eye movements can be indicative of cranial nerve dysfunction, stroke, or vestibular disorders.

[0084] The gaze position may refer to the focal point within the visual field where the subject’s eyes are directed at any given time. The input layer assigns tags to the horizontal and vertical coordinates of the gaze position relative to the presented stimulus. Tagged data for gaze position can be processed for assessing reflexive gaze stabilization and voluntary eye control. Deviations from expected gaze positions can be indicative of potential ocular motor deficits, vestibular dysfunctions, or cognitive impairments. For example, an inability to maintain a steady gaze on a moving stimulus may generate tagged data indicating impaired ocular motor function.

[0085] The eyelid dynamics may refer to the movement patterns of the upper and lower eyelids during presentation of visual stimuli such as, for example, during optokinetic reflex (OKR) tests. The input layer can tag parameters such as palpebral fissure area (PFA), eyelid closure speed, and reflexive eyelid responses. Tagged data for eyelid dynamics includes measurements of lid asymmetry, delayed eyelid responses, or excessive PFA changes. For example, Von Graefe’s sign, indicative of hyperthyroidism, may be related to delayed upper eyelid movement during a downward gaze. Tagged data for eyelid dynamics can be indicative of conditions such as, but not limited to, cranial nerve dysfunction, stroke, or hyperthyroidism, for example.

[0086] The input layer provides an interface between the data pre-processing pipeline and the neural network. The input layer may transform the pre-processed eye movement data by converting parameters having numerical values on a continuous scale, such as amplitude, frequency, and velocity, into classified data with normalized numerical values. The input layer may assign discrete numerical or symbolic identifiers to an eye movement data point, correlating the datum with a tagged classification or category that defined during the training process. The transformation process includes mapping a tagged reflex type to a distinct category or class for consistent data organization before processing by subsequent layers. Tagged classifications include predefined labels or categories assigned to specific data points in the pre-processed eye movement data, such as reflex types (e.g., optokinetic nystagmus or optokinetic after-nystagmus) or eye movementattributes (e.g., saccade amplitude or nystagmus frequency), to enable structured analysis and pattern recognition by the downstream layers or nodes of the neural network.

[0087] In at least one embodiment, the tagged classifications or categories may relate to a condition the classification model is configured / trained to assess. For example, tagged classifications for eye movement data might include “pursuit gain” and “interruptions,” which may be indicative of conditions such as dementia. The input layer may identify the data values associated to the tagged classifications, such as “pursuit gain” and “interruptions,” from the subject’s eye movement data. The input layer may then assign a numerical identifier by linking classification(s) to the corresponding values in the subject’s data. For example, if the subject’s eye movement data includes measurements for “pursuit gain” and “interruptions,” the input layer assigns a numerical identifier such as “1” to the category “pursuit gain” and “2” to the category “interruptions.” The subject’s data values associated with the classification or category, such as a pursuit gain of 0.75 or an interruption frequency of 3 events per second, are grouped under their respective identifiers (“1” or “2”) to form the classified data.

[0088] The normalization performed by the input layer provides that the eye movement data is scaled within a consistent range suitable for processing by subsequent layers in the classification model. If the pursuit gain is measured as a raw value, the input layer may adjust the raw values to a standardized numerical range that correspond to with the neural network’s desired input format.

[0089] In at least one embodiment, the input layer is configured to perform feature extraction processes configured to generate feature data relevant to specific condition(s) examined by the classification model. A classification model can be configured to test a particular condition such as, but not limited to, stroke, dementia, vestibular disorders, fatigue, intoxication, and cranial nerve palsies, for example, by training the classification model on training data for the conditions as well as normative data (e.g., data from individuals not having one or more of those conditions as training can be performed for the model(s) separately). The input layer can isolate and organize features from the classified data that are relevant to that condition, to generate feature data.

[0090] The feature extraction process includes analyzing the classified data to detect parameters that correspond with the labels or testing requirements of the specificclassification model. For example, in a classification model configured for dementia-related analysis, which processes parameters such as pursuit gain, saccadic delay, and / or interruption frequency, the input layer may isolate and extract data corresponding to pursuit gain and / or interruption parameters from the classified data to generate feature data.

[0091] The feature data generated by the input layer includes a structured set of extracted parameters and their associated classified data. For example, the input layer may generate a normalized pursuit gain value of 0.75 and an interruption frequency of three events per second, along a categorical tag indicating “interruptions” relevant to dementia testing. By isolating and prioritizing features specific to a condition, the input layer delivers relevant inputs in form of feature data to the inference layer in the classification model.

[0092] The feature data includes structured output generated by the input layer after normalizing and extracting relevant attributes from the pre-processed eye movement data. The feature data is a combination of classified data and high-level extracted parameters that are selected to meet the requirements of the classification model. The feature data includes normalized numerical values for continuous parameters, such as but not limited to any combination of amplitude, frequency, velocity, symmetry, and directionality of eye movements, for example, and / or categorical representations of tagged classifications, such as but not limited to, for example, reflex types or movement attributes. For example, for eye movement data captured from a subject for a dementia-related classification model, the feature data may include a normalized pursuit gain value of 0.75, representing the ratio of eye velocity to stimulus velocity over a defined interval, and an interruption frequency of three events per second, which reflects inconsistencies in smooth pursuit. Additionally, the feature data includes categorical tags, such as “1” for pursuit gain and “2” for interruptions, for example.

[0093] In at least one embodiment, the classification model includes an inference layer. The inference layer is configured to process the feature data received from the input layer and determine whether the feature data indicates the presence of the specific condition examined by the classification model. The inference layer may operate using a neural network trained on labeled data associated with the condition. The labeled training dataincludes feature sets tagged with their corresponding classifications, such as “presence of condition” or “absence of condition.” Using the training, the inference layer develops a set of decision rules via weights that correlate specific patterns or tags in the feature data with their associated outcomes.

[0094] The inference layer (also referred to as hidden layer) evaluates the feature data against the trained neural network model. The nodes in the inference layer of the trained classification model applies the learned weights to the input parameters (e.g. feature values) or tags, such as normalized pursuit gain or interruption frequency, to calculate intermediate values at various node(s) in the network. For instance, a high pursuit gain coupled with frequent interruptions may trigger specific activations in the hidden nodes, indicating a likelihood of dementia. The node(s) may process the inputs through activation functions to produce intermediate outputs that represent the contribution of the feature(s) to the final classification.

[0095] The inference layer can also process the intermediate outputs to aggregate the contributions of all features or tagged data in the feature data. The aggregation may occur across multiple nodes within the interference layer. The inference layer can refine the analysis based on the patterns learned during training. For example, the trained classification model may combine the normalized pursuit gain value with interruption frequency to determine a composite score that represents the likelihood of the condition for which the classification model was trained. The inference layer of the neural network outputs evaluation data indicating the extent to which the feature data matches the patterns associated with the examined condition. The evaluation data may include a classification score or probability.

[0096] The classification score from the inference layer may represent an assessment of the feature data. For example, in a dementia-related classification model, the inference layer may generate a score of 0.85, indicating an 85% likelihood of the condition being present based on the analyzed features. The classification model may also be implemented such that it generates evaluation data that can also, or alternatively, include categorical results, such as “Condition Detected” or “Condition Not Detected,” along with confidence metrics. For example, confidence scores may be provided by incorporating the softmax function, for example.

[0097] The classification model includes an output layer. The output layer is configured to receive evaluation data from the inference layer and generate the assessment data of the classification model. The evaluation data received from the inference layer may include a classification score or probability, and intermediate outputs representing the aggregated contributions of the analyzed feature data. The output layer processes the evaluation data to produce an interpretable result in form of assessment data. The assessment data may be generated based on a mapping of the analyzed feature patterns from the feature data to specific classification categories relevant to the condition examined by the classification model.

[0098] In at least one embodiment, the output layer may be configured to apply decision thresholds to the classification score received from the inference layer. For instance, a classification score exceeding a predetermined threshold, such as but not limited to 0.85, for example, may correspond to a positive detection of the condition, while a score below the threshold may indicate the absence of the condition. The decision thresholds can be predefined during the training phase of the classification model or directly received as user data.

[0099] In at least one embodiment, upon determining the classification category, the output layer generates the assessment data to include a structured result that includes the classification category, associated confidence metrics, and any additional data relevant to the condition-specific analysis and decision-support. For example, in a dementia-related classification model, the output layer may generate assessment data indicating “Condition Detected” with a determined confidence score, such as 85%, for example, accompanied by a breakdown of the informative contributing features, such as a pursuit gain value and interruption frequency values.

[0100] In at least one embodiment, the output layer may also format the results for integration with external systems or display devices. The results may be encoded in a standardized format, such as JSON or XML, to provide interoperability with decisionsupport tools, reporting systems, or cloud-based platforms. For example, the output layer can transmit the results to a connected device, such as an administrator’s workstation or a remote server, for visualization or further analysis. Alternatively, in at least oneembodiment, one or more post-processing / output software modules may be used to perform these function(s).

[0101] In at least one embodiment, the assessment data generated by the output layer may include pattern correlation data. The pattern correlation data can include longitudinal tracking of eye movement parameters, comparative analysis against baseline measurements, or identification of co-occurring abnormalities. For example, in a vestibular disorder classification model, the output layer may identify patterns indicative of multiple overlapping conditions, such as combining vestibular hypofunction with nystagmus asymmetry, and generate annotations explaining these findings. The pattern correlation data may also include predictive analytics derived from trends observed in the feature data. For instance, the output layer may analyze gradual changes in pursuit gain or interruption frequency to predict potential progression of conditions like dementia or Parkinson’s disease. In at least one embodiment, the output layer may include one or more regression models and / or perform trend analysis techniques to generate the pattern correlation data such that it may provide insights into how the subject’s condition might evolve over time.

[0102] In at least one embodiment, the pattern correlation data may be presented with visualizations, such as trend graphs or comparative tables. The pattern correlation data can be encoded in interoperable formats. In embodiments in which the output layer’s functionality is extended beyond detection, the output layer provides additional capabilities to the classification performed by the trained classification model for a particular condition.

[0103] In at least one embodiment, the output layer is further configured to generate administration data. The administration data provides actionable decision-support outputs regarding treatment options and / or recommended next steps based on the condition detected by the classification model. Upon determining a positive classification, the output layer may be configured to reference a pre-defined database including condition-specific guidelines, or decision-support system connected to the trained classification model. The database may be programmed with evidence-based guidelines and treatment protocols associated with the detected condition for decision-support purpose. For instance, in a dementia-related classification model, the administration data may be generated toinclude recommendations for cognitive rehabilitation exercises, pharmacological interventions, or referrals to a neurologist based on the detected severity level derived from the classification score.

[0104] In at least one embodiment, the administration data may be generated to also include time-sensitive actions or follow-up steps depending on the condition. For example, for conditions requiring immediate medical attention, such as stroke-related abnormalities detected through specific eye movement patterns, the administration data may include an alert suggesting emergency intervention and provide contact details for the nearest healthcare facility. The output layer integrates the administration data into the assessment data pipeline.

[0105] In at least one embodiment, the output layer may be configured to generate realtime feedback including alerts, progress indicators to show the stage of data processing, and preliminary data patterns providing initial trends or anomalies detected during analysis. The real-time feedback can be displayed on connected devices, such as a workstation or a handheld device, and may provide actionable insights and / or recommendations based on intermediate results from the classification model.

[0106] Reference is first made to FIG. 1 , which illustrates a block diagram of an example embodiment of a pattern assessment and administration system 100. The system 100 includes a network 104 that connects multiple components of the system 100. The system further includes a user device 102 (also referred to as user apparatus 102), network 104, a multi-model architecture 106, an administrator system 108, an external data storage 110, and an external system 112. The multi-model architecture 106 further includes a plurality of Al models, i.e., trained classification models, some of which are illustrated as Model A 120, Model B 122, and Model Selection Module 124.

[0107] The system representation shown in FIG. 1 is provided as an example embodiment. There may be variations in the combination or number of such components, and in some cases, a single device may provide the functions of multiple components.

[0108] The user device 102 (also referred at as user apparatus 102) can be a portable device configured to present optokinetic stimuli and capture eye movement data. The device 102 is configured to operate in diverse environments, including remote locations, using cloud connectivity for data storage and at least some processing while some dataprocessing can also be done locally at the device 102. In an embodiment, the device 102 may be implemented as a medical-grade device or consumer-grade hardware such as a smartphone, tablet, or laptop equipped with a camera. The device 102 is configured to capture eye movements of a subject and generate eye movement data. The eye movement data may include parameters such as amplitude, frequency, velocity, and symmetry of eye movements, as well as auxiliary features like gaze direction and eyelid dynamics. The eye movement data can be transmitted to the other component(s) of system 100 via the network 104.

[0109] The user device 102 may include a display configured to present optokinetic reflex (OKR) tests or other visual stimuli to elicit specific eye movement patterns. While the device is shown to be a computer in FIG. 1, it may be implemented using different hardware and include a headset or goggles for presentation of test signals / stimulus signals to a subject during the OKR testing. The test stimuli may include various visual patterns such as alternating stripes, moving dots, or rotating fields, which may trigger reflexive responses like optokinetic nystagmus (OKN) or smooth pursuit. The display may also be used to present custom visual tasks or assessments, such as evaluating gaze stabilization, fixation, or saccadic movements. The display may show standardized stimuli such as alternating light and dark stripes (an example is shown in FIG. 5), which elicit optokinetic nystagmus (OKN). The optokinetic nystagmus includes a repetitive, involuntary eye movement pattern characterized by alternating smooth pursuit movements in the direction of a moving stimulus and rapid saccadic movements that reset the gaze. The device 102 can vary the width, speed, and color of the lines or provide regular black and white striped patterns to ensure reliable reflex elicitation. The display parameters, including stimulus speed, size, and distance. The display device may include smartphone screens, tablet screens, laptop or desktop monitors, television screens, virtual reality headsets, augmented reality headsets, wearable visors, projector-based displays, head-mounted displays, specialized medical displays, heads-up displays, smart glasses, or certain combinations thereof.

[0110] The device 102 may include an imaging device such as an integrated or connected camera for capturing the subject’s eye movements during the presentation of stimuli to generate the eye movement data. In an embodiment, the imaging deviceincludes high-resolution eye-tracking cameras to record detailed eye movement patterns. Infrared or near-infrared sensors may be included in the imaging device to improve accuracy by reducing sensitivity to ambient light. Further, for assessments such as optokinetic after-nystagmus (OKAN), which use total darkness, the camera system may include infrared sensors to record the subject’s sustained eye movement response after the cessation of the visual stimulus. The imaging device may include standard cameras, infrared cameras, high-speed cameras, depth cameras, webcams, specialized medical imaging devices, wearable cameras, thermal cameras, eye-tracking sensors, or certain combinations thereof.

[0111] In an embodiment, the device 102 may be implemented as virtual reality (VR) glasses. The VR glasses may include embedded high-resolution cameras and displays to provide immersive stimuli and eye-tracking capabilities. The VR glasses may be paired with cloud-based systems or other components of the system 100 via the network 102.

[0112] In an embodiment, the user apparatus 102 may include a visor to standardize testing conditions. The visor, in combination with infrared sensors, may provide consistent distance between the display and the subject’s pupils for alignment of presentation of visual stimuli as well as a consistent distance between the camera and the subject’s eyelids. The visor may improve user device 102 alignment for accurate data capture. The visor may be implemented for assessments requiring full-field stimulation, such as OKAN.

[0113] The user device 102 may include adjustable lighting control systems to maintain consistent and desired lighting during testing. Diffused light sources may be used to minimize glare and shadows.

[0114] In at least one embodiment, the user device 102 may include mounting equipment, such as adjustable mounts for cameras or wearables like visors or glasses with embedded cameras. The mounting equipment is provided for stability during data capture.

[0115] The network 104 is configured to interconnect the components of the system 100, and provide data exchange among the user device 102, the multi-model architecture 106, the administrator system 108, the external data storage 110, and the external system 112. The network 104 may provide for the transfer of eye movement data, classification results, and additional insights generated by the system 100. The network 104 may alsosupport secure access to cloud-based platforms for data storage, processing, and remote monitoring.

[0116] The network 104 can include various configurations, such as a local area network (LAN), a wide area network (WAN), or a combination of both, depending on the deployment environment. For remote and portable applications, the network 104 may implement wireless communication protocols such as Wi-Fi, cellular networks (e.g., 4G or 5G), or Bluetooth. In at least one embodiment, the network 104 can implement wired connections, such as Ethernet. The network 104 may also implement encryption protocols and firewalls to provide the confidentiality of sensitive medical data during transmission.

[0117] The multi-model architecture 106 is a processing component implemented using one or more neural network models that execute on one or more processors so that the one or more processors are configured to process eye movement data received from the user device 102. The one or more processors are also configured for generating one or visual stimuli that are then output on the display for viewing by the subject. The architecture 106 includes a plurality of classification models, such as Model A 120, Model B 122, and Model Selection Module 124. The classification model(s) are configured to determine (and optionally analyze) neurological and / or physiological conditions from patterns identified in the eye movement data. In at least one embodiment, the condition may refer to non-medical states or behaviors, such as levels of alertness, cognitive workload, or fatigue, determined during the training process based on correlations between labelled eye movement data patterns and the behavioral states. The architecture generates assessment data that maps the analyzed features to the identified conditions and optionally administration data that provides actionable next steps or treatment recommendations. The multi-model architecture 106 can be implemented on a server.

[0118] The implementation of the multi-model architecture 106 may include distinct artificial intelligence-based classification models 120, 122, 124 that operate independently or in combination with one another. The classification models are trained Al models that can be configured / train to detect one or more conditions and perform analysis which may result in certain recommendations. For example, one classification model is configured to identify neurodegenerative disorders, while another classificationmodel is configured to identify vestibular impairments. The modular architecture 106 provides for integration or removal of classification model(s).

[0119] The multi-model architecture 106 may include features such as dynamic resource allocation, cloud integration, and / or real-time data processing. Dynamic resource allocation can distribute computational workloads among the classification models, for improved performance (e.g., reduced processing time). Cloud integration can offload processing to distributed servers to improve scalability.

[0120] The administrator device 108 is configured to manage the operations of the system 100. The administrator system 108 may configure or monitor components such as the multi-model architecture 106, user device 102, and the network 104. In at least one embodiment, the administrator device 108 is configured to manage system settings, such as threshold values for classification models, condition determination protocols, and data privacy controls. The administrator device 108 can receive, display, and analyze assessment data and administration data generated by the multi-model architecture 106 to healthcare professionals or system operators.

[0121] The administrator device 108 may be implemented as a desktop workstation, a laptop, a smartphone device, a tablet computer, or a web-based platform accessible through secure portals. The administrator device 108 may provide user interfaces for realtime visualization of assessment results and graphical representations of the assessment data. The administrator device 108 can also communicate with external systems 112 or databases 110 to store or retrieve subject records, treatment guidelines, and / or research data.

[0122] The external data storage 110 may be used (e.g., configured) to store data generated, processed, or used by the system 100. The storage 110 may include cloudbased servers, local databases, or hybrid solutions. The external data storage 110 can store pre-processed eye movement data and feature data generated by the input layer of a given trained classification model. The external data storage 110 can store assessment data and administration data produced by the multi-model architecture 106. The storage 110 may also include historical patient records, clinical guidelines, assessment protocols, tagged / labelled training datasets for the classification models, and / or configuration filesfor system settings. The implementation may include encryption protocols and access controls.

[0123] The external system 112 can include any third-party platform, device, or networked resource that interacts with the components of the system 100. The external system 112 may include healthcare information systems, such as electronic health records (EHR) platforms, which integrate patient data from the system 100 into broader clinical workflows. For example, assessment data and administration data generated by the classification models in the multi-model architecture 106 can be transmitted to an EHR system for storage, further review, or use in treatment planning. The external system 112 may also include decision-support tools, such as imaging platforms or lab testing systems to supplement the system’s findings.

[0124] Reference is now made to FIG. 2, which illustrates a block diagram of an example embodiment of a classification model selection system 200, which may be employed by the system 100. The system 200 includes a Model Selection Module 204, a Multi-model Architecture 206, and a Selected Classification Model 210.

[0125] In at least one embodiment, the system 200 is configured to determine and select an appropriate classification model from a multi-model architecture 206 based on the input data 202. The system 200 includes a model selection module 204, which receives eye movement data and optionally visual stimulus data (not shown) as input data 202. The visual stimulus data may be associated with the optokinetic reflex (OKR) tests. The model selection module 204 processes the received input data using a set of selection parameters and generates a selection signal 205. The selection signal 205 is used in the multi-model architecture 206 to activate the selected classification model 210 for further processing.

[0126] The model selection module 204 receives eye movement data 202. The eye movement data 202 which may include data signals such as, but not limited to, any combination of amplitude, frequency, velocity, pursuit gain, interruptions, symmetry, and directionality of eye movements. Additionally, the eye movement data 202 may include auxiliary data, such as gaze position and / or eyelid dynamics, as part of the feature data provided by the user device. The module 204 is also configured to receive stimulus data that includes the type, characteristics, and configuration of the visual stimuli displayed tothe subject during the OKR test. For example, the stimulus data may include information about the width, speed, color, and direction of the stripes used to elicit optokinetic nystagmus. The stimulus data may include testing conditions, such as the presence or absence of ambient light.

[0127] The model selection module 204 processes the input data 202 to evaluate a set of selection parameters. The selection parameters may include the type of eye movement data received, such as whether the data represents saccadic movements, smooth pursuit, or reflexive gaze stabilization. Additional parameters may include tagged features in the training dataset, such as pursuit gain and interruption frequency, which can provide indication about the potential condition being identified and optionally analyzed. The characteristics of the stimulus data, such as its speed and direction, may also be used as a parameter to assess the condition. User input based on factors such as a suspected condition or prior medical history, may also be included as a selection parameter.

[0128] Based on the evaluated parameters, the model selection module 204 performs a probabilistic analysis to identify the most appropriate classification model for processing the data. The module 204 can be implemented as an artificial intelligence-based probabilistic system trained on labeled datasets of eye movement patterns and associated assessment outcomes. The probabilistic analysis may include assigning weights to different selection parameters and calculating a likelihood score for the trained classification model(s) indicating which trained classification model should be used for identifying the condition and optionally performing further analysis. For example, if the eye movement data indicates frequent interruptions in smooth pursuit and the stimulus data corresponds to high-speed stripes, the module 205 may determine a high probability that the condition being analyzed is related to neurodegenerative disorders, such as dementia. In response, the module 205 may generate a selection signal 205 to indicate that the trained classification model associated to dementia assessment and administration should be used to further process the eye movement data.

[0129] The model selection module 204 generates a selection signal 205 based on the output of the probabilistic analysis. The selection signal 205 identifies the trained classification model within the multi-model architecture 206 to be activated for processing the input data. For instance, the selection signal 204 may direct the activation of trainedClassification Model 1 if the analyzed parameters correspond to the conditions evaluated by that model. Alternatively, trained Classification Model 2 can be selected if the selection parameters associated to the second model are identified.

[0130] Alternatively, in at least one embodiment, the user conducting the OKN testing may provide input to select one or more trained classification models for processing the eye movement data. The user input is configured to select one or more of the trained classification models for processing the eye movement data for targeted analysis based on the user preferences, the subject’s history, and / or prior assessments. For example, the selected classification model(s) may analyze the current eye movement data to evaluate whether the identified condition has improved or worsened by comparing the test results with results obtained during previous testing sessions conducted at different points in time.

[0131] Reference is now made to FIG. 3, which illustrates a block diagram of an example embodiment of a trained classification model 300. The trained classification model 300 includes an Input Layer 304, a representation of Feature Data 306, an Inference Layer 308, and an Output Layer 310. The trained classification model 300 may correspond to the selected classification model 210 described in FIG. 2.

[0132] The input layer 304 of the trained classification model 300 is configured to receive and process eye movement data captured during optokinetic reflex (OKR) tests or other stimuli by segregating and tagging the data points based on predefined classifications. The data point(s) in the eye movement data correspond to the measurable parameters, such as any combination of amplitude, frequency, velocity, symmetry, and directionality. The parameters are assigned tags that categorize the parameters into distinct categories based on the nature of the measured eye movement, including nystagmus, saccades, smooth pursuit, look-to-stare ratios, and / or palpebral fissure area (PFA). The tagged classifications enable the input layer 304 to organize the data in a structured format.

[0133] The input layer 304 is configured to categorize and tag eye movements captured from the imaging device. The tagging process provides for labelling the captured eye movement.

[0134] The tagged classifications in the eye movement data may include nystagmus-related parameters, such as any combination of amplitude, frequency, direction, andsymmetry. For example, the input layer 304 may process the eye movement data related to the amplitude and frequency of optokinetic nystagmus (OKN) triggered by a moving striped stimulus and assign tags to the eye movement data.

[0135] The tagged data, also known as classified data, may include eye movement data in normalized values and labeled with predefined classifications based on measurable parameters. For example, a recorded eye movement exhibiting a simultaneous and rapid shift of gaze is tagged as a “saccade,” while slower, tracking movements are tagged as “smooth pursuit”. The input layer 304 can assign tags to the data points by a tagging algorithm trained on labeled datasets during the training phase of the classification model. The training datasets may correlate eye movement data with predefined classifications. The input layer 304 can implement the tagging algorithm to identify and label the eye movement data. In at least one embodiment, adaptive learning mechanisms may be implemented within the input layer 304 to further refine the tagging process. The adaptive learning can be provided by receiving user feedback to verify or adjust assigned tags on new data and dynamically updating the tagging rules based on the user feedback.

[0136] Saccadic movements are another category of tagged classifications. The input layer 304 may tag parameters such as speed, accuracy, and reset delay of saccades. The saccades are rapid, discrete movements that redirect the gaze. The saccadic tags can be relevant for analyzing gaze stabilization and assessing conditions such as neurodegenerative disorders or traumatic brain injuries. For instance, delayed or erratic saccades may be indicative of neurological dysfunctions. The tagging and normalized measurements of the saccadic features allows the classification model 300 to process condition-specific data points during the feature extraction and inference process.

[0137] The tagged classifications may include smooth pursuit movements. The smooth pursuit movements refer to the eyes tracking a moving object. The smooth pursuit movements can be tagged with classifications such as pursuit gain and interruptions. The pursuit gain represents the ratio of eye movement velocity to stimulus velocity. The interruptions represent breaks or inconsistencies in the smooth tracking. The tagged classifications provide for identifying conditions such as dementia, where smooth pursuit is impaired. The input layer 304 may provide for isolating and tagging the smooth pursuit parameters.

[0138] The tagged classifications may include look-to-stare ratio to distinguish between reflexive (subcortical) and voluntary (cortical) eye movements. Reflexive movements are automatic and triggered by subcortical pathways. The voluntary movements are triggered by cortical involvement. The look-to-stare ratio may provide data for assessing cortical and subcortical brain functions. The look-to-stare ratio can indicate cognitive decline or brainstem dysfunctions. The input layer 304 assigns tags to differentiate between the movement types.

[0139] For example, in at least one embodiment, the input layer 304 may be configured to assign tags related to the Look-to-Stare Ratio derived from the relationship between reflexive and voluntary eye movements during vertical optokinetic stimulation (OKS). The Look-to-Stare Ratio captures data on palpebral area changes and eye movement patterns to capture differentiation between cortical-driven “look” responses and subcortical-driven “stare” responses. The tagged data may include measurements of eyelid position changes (palpebral aperture), and saccadic resets recorded during a controlled vertical stripe pattern test that elicits optokinetic nystagmus (OKN). The input layer 304 processes the captured data to identify waveform patterns generated by palpebral area changes as the eyes track moving stripes. Look responses, characterized by smoother, higher-amplitude waveforms, reflect active, cortical-driven tracking and involve frontal eye fields and parietal lobes. In contrast, stare responses, which exhibit lower-amplitude, irregular waveforms, rely on subcortical pathways such as the brainstem and superior colliculus. The Look-to-Stare Ratio is calculated based on waveform amplitude, frequency, and stability, providing a quantitative measure of cortical engagement. The Look-to-Stare Ratio can provide an early marker for conditions like dementia, where cortical degeneration impairs voluntary tracking before more severe cognitive deficits manifest. A lower Look-to-Stare Ratio, determined when compared to a threshold, may indicate reduced cortical function, enabling non-invasive, quantitative assessment of neurological conditions. The tagging of palpebral area changes derived from eyelid movement as a proxy for eye movement, particularly during OKN testing, is advantageous for determining the Look-to-Stare Ratio. The process reduces the reliance on more complex or intrusive equipment, making the valuable data accessible for early detection of neurodegenerative diseases.

[0140] In at least one embodiment, the input layer 304 may also be configured or trained to assign tags related to palpebral fissure area (PFA). The palpebral fissure measurements capture changes in eyelid separation during the presentation of a vertical OKR stimulus to the subject. The PFA includes the vertical opening between the upper and lower eyelids, which may change during visual testing, and is therefore measured over time during eye movement tests. The PFA is generally represented as the area enclosed by the eyelids when the eye is open. The PFA measurement includes capturing the distance between the upper and lower eyelids at various time points in response to optokinetic stimuli (e.g., moving visual patterns). The PFA measurement can be performed by capturing real-time video of the eye from an imaging device, detecting the edges of the upper and lower eyelids through segmentation algorithms, calculating the vertical distance between the edges at various time points, and mapping the measurements over time to generate a waveform representing eyelid separation dynamics. It has been determined by the inventors that the amount / type of PFA changes over time during visual testing can relate to various conditions such as, but not limited to, a stroke, cranial nerve palsy, and hyperthyroidism, for example. Symmetrical PFA may refer to the state of similar vertical distance between the upper and lower eyelids for both eyes during a stimulus. Asymmetrical PFA i.e., when one eyelid of one eye moves differently from the eyelid of the other eye, can be indicative of cranial nerve palsy (one eyelid droops or moves slower), stroke (unilateral loss of reflexive eyelid movement), and hyperthyroidism (one eyelid may lag behind the other).

[0141] The palpebral tags can include a variety of data parameters related to PFA measurements, which are categorized to identify a variety of conditions affecting neurological, ophthalmological, and endocrine systems. The palpebral fissure tags can be processed to detect vertical reflex abnormalities and conditions such as cranial nerve dysfunction or hyperthyroidism. For example, irregular PFA changes during vertical OKR may indicate a loss of vertical gaze stabilization or eyelid control. The additional PFA tagged data may include parameters such as PFA Symmetry, measuring differences in PFA between both eyes to detect asymmetry indicative of cranial nerve palsy or stroke. PFA tags capturing asymmetric eyelid movements may indicate stroke-related abnormalities, suggesting potential damage to cranial nerves such as CN III. PFA tagsassociated with conditions like cranial nerve palsy can include ptosis or drooping eyelids, reflecting dysfunction in cranial nerves such as CN III, IV, or VI. For concussion or mild traumatic brain injury (mTBI), PFA tags recording variability in eyelid responses or delayed recovery times provide indicators of impaired reflex control linked to brain injury. The PFA Response Time tag tracks the time for eyelids to respond to a stimulus to identify reflex delays linked to neurological dysfunction. The PFA Amplitude Ratio tag calculates the ratio of maximum PFA change to baseline PFA for detecting conditions such as hyperthyroidism or stroke. For endocrine disorders, the PFA tags may include lid lag during downward gaze and increased PFA amplitude, which are indicative of hyperthyroidism, particularly Graves' disease. Such measurements detect impairments in eyelid reflexes linked to thyroid overactivity. Additionally, conditions such as myasthenia gravis may present with fluctuating PFA values that provide markers for autoimmune disorders affecting eyelid muscles. For vestibular and degenerative disorders, the PFA tags may include asymmetry during optokinetic reflex (OKR) tests, which can disrupt oculomotor reflex pathways in vestibular disorders. Reduced PFA variability and slower reflexive responses are associated with Parkinson’s disease, indicating neurodegenerative motor dysfunction. For conditions like Bell’s palsy, asymmetric PFA due to facial nerve paralysis reflects weakness in CN VII, leading to drooping or non-responsive eyelids. The PFA tags capturing reduced variability and slower reflexive responses may be indicative of Parkinson's disease. For conditions such as Bell’s palsy, PFA tags include asymmetric eyelid movements, drooping or non-responsive eyelids, generally caused by facial nerve paralysis can reflect CN VII weakness. Accordingly, depending on which one or more conditions might be tested by a given embodiment, the tags described above related to these one or more conditions may be used by the trained classifier.

[0142] In at least one embodiment, the input layer performs a normalization process after tagging the eye movement data to generate classified data. The subject’s measurements on tagged parameter(s) can be converted into a normalized numerical value for consistency and compatibility with the classification model’s input specifications. For example, a pursuit gain value of 0.75 may be normalized to fall within a range of 0 to 1 ,and interruptions may be encoded as discrete events per second. The normalization provides for standardizing the tagged data points.

[0143] The input layer 304 is configured to process eye movement data or parameters received from the user device by transforming the parameters with continuous numerical values into classified data. The eye movement data may include continuous measurements such as any combination of amplitude, frequency, velocity, symmetry, and directionality of movements. The eye movement data point(s) are associated with parameters that describe the subject’s response to a given stimulus. The input layer 304 converts the continuous values into normalized numerical values compatible with the selected classification model 300. For instance, if the amplitude of an eye movement is recorded as 15 millimeters, the input layer 304 normalizes the raw data value to fall within a predefined numerical range, such as 0 to 1 , to provide consistency in data format.

[0144] The tagging process performed by the input layer 304 may assign discrete numerical or symbolic identifiers to an eye movement or reflex present in the eye movement data. Tagged classifications may include predefined labels such as “pursuit gain,” “saccade speed,” or “nystagmus frequency.” For example, a measured pursuit gain of 0.75 may be assigned the tag “1.” Similarly, an interruption frequency of three events per second can be tagged as “2.” The input layer 304 can group the corresponding normalized values with their respective tags to create a structured dataset. In at least one embodiment, the classified data may include the normalized and standardized version of tagged eye movement data. The input layer 304 processes tagged eye movement data by scaling raw measurements to standardized numerical ranges. The normalized and classified data can provide compatibility with the classification model’s input data requirements.

[0145] In an embodiment, the classified data generated by the input layer 304 includes the subject's normalized eye movement data and the temporal or spatial association with the corresponding stimulus that induced the eye movement. For example, if a stimulus having a high-speed horizontal stripe pattern triggered a specific reflex, such as optokinetic nystagmus, the classified data includes the normalized frequency and amplitude values of the eye movement and the metadata describing the stimulus characteristics. For example, classified data can represent entries such as “Tag 1 : PursuitGain, Value: 0.75, Stimulus: Horizontal Stripes with 1.5 Width, Timestamp: T1.” The association provides for enriching the classified data with a representation of both the eye movements and the triggering conditions.

[0146] In at least one embodiment, the input layer 304 can perform additional processes to improve data organization. Tagged classifications may be mapped to categories or classes based on the specific condition being analyzed. For example, during an OKR test for dementia detection, tags such as “pursuit gain” and “interruptions” are prioritized and grouped under a class related to smooth pursuit analysis. The input layer 304 extracts the tagged features from the raw data, normalizes the tagged features, and assigns the normalized data to their respective classifications. The raw value(s) from the eye movement data are scaled to reduce variations caused by measurement units or testing conditions. For example, if velocity is recorded in millimeters per second, the input layer 304 standardizes this parameter to a dimensionless numerical scale.

[0147] In at least one embodiment, the input layer is configured to implement feature extraction process on the classified data within the input layer. The input layer isolates the tagged values relevant to the condition being analyzed by the classification model to generate categories of feature data 306. For example, when evaluating for dementia, the input layer 304 extracts tags such as pursuit gain, interruption frequency, and saccadic reset delay while excluding parameters irrelevant to dementia assessment. The extraction process reduces the computational burden on the inference layer 308 by minimizing the volume of data that the inference layer 308 processes. By combining tagging, normalization, and feature extraction, the input layer 304 delivers structured and condition-specific feature data to the subsequent layers including the inference layer 308 and the output layer 310.

[0148] The feature data 306 includes a subset of classified data extracted by the input layer 304 based on the classified data’s relevance to the specific condition being evaluated by the selected classification model 300. For instance, if the trained classification model is configured to detect neurodegenerative disorders, the model processes classified data associated with pursuit gain, interruption frequency, and / or saccadic delay. The feature extraction process may include extracting the data values on pursuit gain, interruption frequency, and / or saccadic delay, and transmitting the datavalues to the inference layer 308 for further processing. The feature extraction may exclude irrelevant parameters thereby reducing data volume and improving computational efficiency for the trained classification model.

[0149] The determination of tagged classifications relevant for a condition and trained classification model may be based on preset assessment tags assigned during the training phase of the classification model. The preset assessment tags can be derived from clinical research, guidelines, and / or labeled training datasets that correlate specific eye movement parameters with particular conditions. For example, for a dementia-related trained classification model, tagged classifications such as pursuit gain, interruption frequency, and saccadic reset delay are identified as relevant. The tags can be predefined in the model’s configuration and perform as a filter. The input layer 304 provides that the relevant tagged feature(s) are extracted and analyzed during the feature extraction process.

[0150] In addition to preset assessment tags, the determination of relevant tagged classifications may also be determined by adaptive learning algorithms within the trained classification model. The classification models, during the training phase, analyze historical data and outcomes to refine the selection of tags based on their significance to the condition(s). For instance, a trained classification model may be trained to identify new correlations between previously untagged features and specific conditions through ongoing training on updated datasets. User input or clinician-specified criteria may also configure the preset assessment tags. For example, a clinician may prioritize certain parameters or tags based on a subject’s medical history or symptoms.

[0151] During the training phase, the classification model can receive labeled datasets where tagged data point(s) are associated with a known outcome, such as “presence of condition” or “absence of condition.” Positive reinforcement is applied when the model correctly identifies eye movement data or features associated with the desired outcome by strengthening the weights assigned to those features. Conversely, negative reinforcement is applied when the model produces an incorrect classification by implementing adjustments to the weights or decision rules. The iterative process can refine the classification model’s ability to recognize relevant patterns in the tagged data for improved accuracy.

[0152] In at least one embodiment, the input layer 304 processes palpebral fissure area (PFA) measurements by tagging, normalizing, and extracting data relevant to vertical optokinetic reflex (OKR) tests or similar stimuli exposures. During vertical OKR stimulation, the cameras record the subject’s eyelid region to capture continuous changes in the vertical distance between the upper and lower eyelids over time. The input layer 304 can apply segmentation computations to identify the edges of the upper and lower eyelids and calculate the PFA at the time frame(s). The tagged classifications for PFA measurements generated by the input layer 304 may include parameters such as eyelid separation amplitude, symmetry, and / or reflexive responsiveness. Symmetry includes the equal and coordinated movement of both eyes or eyelids of both eyes during the presentation of the visual stimuli. Symmetrical responses are indicative of normal ocular motor function, and any deviations from symmetry may indicate neurological or physiological abnormalities. Reflexive responsiveness includes the involuntary reactions of the eyes and eyelids to visual stimuli, such as those elicited during optokinetic reflex (OKR) tests. The reflexive reactions include synchronized eyelid movements, such as appropriate opening and closing in response to the presented stimulus. Impaired reflexive responsiveness, such as delayed, reduced, or absent movements, may indicate underlying neurological or physiological dysfunctions. Such conditions can include optokinetic dysreflexia (OKD), characterized by diminished reflexive eye movements, cranial nerve damage leading to eyelid muscle control loss, or external factors such as fatigue or intoxication causing slower responses. Additional conditions include hyperthyroidism presenting as lid lag, where the upper eyelid fails to move in coordination, or stroke, resulting in asymmetric or absent eyelid movement. Accordingly, the various embodiments described herein are implemented to detect the various eye or eyelid movements more associated with a condition being tested in a more effective and accurate manner thereby allowing for early onset of the conditions to be detected as well as monitoring the progression of the conditions overtime.

[0153] For example, delayed or absent PFA changes during vertical OKR are tagged as indicative of reflex loss or dysfunction. The abnormalities in palpebral fissure area (PFA) measurements may correlate with certain conditions such as, but not limited to, stroke, hyperthyroidism (evidenced by lid lag), or cranial nerve palsies.

[0154] In at least one embodiment, the input layer 304 can normalize the tagged PFA measurements. For example, raw PFA values recorded in millimeters are scaled to a standardized numerical range, such as 0 to 1. After normalization, the input layer 304 performs feature extraction to isolate PFA tags relevant to the specific condition associated to the selected classification model 300. For a stroke-related trained classification model, the input layer 304 extracts tags representing asymmetry or delayed PFA changes and exclude parameters irrelevant such as those related to horizontal OKR.

[0155] The input layer 304 is configured to extract specific tagged features from the eye movement data based on the condition being examined by the selected classification model. The measurements may include any combination of amplitude, frequency, velocity, pursuit gain, interruptions, symmetry, and directionality of eye movements. The selected measurements are determined by the trained classification model such that the relevant features are extracted for the assessment.

[0156] For example, in at least one embodiment, when the trained classification model 300 is configured (i.e., trained) for nystagmus detection, the input layer extracts tags related to any combination of amplitude, frequency, direction, and symmetry of eye movement data. Abnormal horizontal or vertical eye movement data during optokinetic reflex (OKR) tests are also tagged for analysis. Vertical stimulation eye movement data leading to horizontal eye movements, or vice versa, is extracted as indicative of nystagmus. The tagged features enable the selected classification model to assess reflex abnormalities or oscillatory eye movement patterns that correspond with conditions such as vestibular disorders or brainstem dysfunction.

[0157] Alternatively, in at least one embodiment, for a classification model configured to detect oculomotor palsy detection, the input layer may be configured to extract tagged features related to asymmetries in eye movement data. Parameters such as gaze direction, movement amplitude, and / or symmetry may be tagged and extracted. Asymmetry in reflexive responses during optokinetic testing can be a marker for unilateral oculomotor palsy. The input layer 304 isolates the parameters and normalizes them for model’s further processing.

[0158] As another example, in at least one embodiment, when the trained classification model 300 is configured for hyperthyroidism and lid disorder detection, the input layer 304extracts the tags associated to the palpebral fissure area (PFA) measurements. For example, the tags associated to changes in eyelid separation during vertical OKR are extracted for further processing. The input layer 304 can tag the instances of lid lag during downward eye movement. The lid lags can relate to Von Graefe’s sign and indicative of hyperthyroidism. Tags associated with irregular PFA changes are extracted for further processing by the classification model configured for hyperthyroidism and lid disorder detection.

[0159] As another example, in at least one embodiment, for a classification model assessing auditory nuclei dysfunction (OKAN test), the input layer 304 extracts tagged data points related to sustained eye movements observed in darkness following the cessation of optokinetic stimuli. The input layer 304 isolates parameters such as movement duration and frequency during the OKAN phase. Data tags associated to abnormalities in the parameters, such as prolonged movements persisting for 8-10 seconds, for example, are extracted for being indicative of cranial nerve VIII (auditory nerve) dysfunction. The extracted features are processed by the trained classification model configured for evaluating vestibular and auditory pathways.

[0160] As another example, in at least one embodiment, for a classification model assessing conditions involving decreased consciousness or reflex loss, the input layer 304 extracts the tagged features such as partial or complete absence of ocular movement in response to optokinetic stimuli. Tagged features such as diminished waveform amplitude or a flat-line graph on the optokinetograph may also be extracted. The features are processed by the trained classification model configured for optokinetic dysreflexia (OKD) or associated neurological dysfunctions. The directionality and symmetry of the eye movements are tagged to identify any abnormalities, such as optokinetic dysreflexia (OKD). The optokinetic dysreflexia (OKD) is characterized by asymmetric or absent reflexive eye movements. The OKD tagged data points may indicate neurological and physiological impairments.

[0161] In at least one embodiment, when the trained classification model is configured (i.e. trained) for stroke detection, the input layer 304 may extract tagged data related to asymmetric or delayed reflexes, nystagmus asymmetry, and saccadic response delays. Tagged data such as amplitude, direction, and / or timing of eye movements may beextracted for further processing by the trained classification model. In at least one embodiment, additionally, the tagged irregularities in PFA during vertical OKR may be extracted to detect loss of reflexive lid movement as a marker of a stroke, and brainstem or cerebellar dysfunctions associated with cerebrovascular conditions.

[0162] As another example, in at least one embodiment, for the trained classification model is configured to assess vestibular disorders, the input layer 304 may be configured to extract tagged features associated with off-axis nystagmus during vertical OKR or abnormal OKAN responses including nystagmus amplitude, frequency, and directionality. Tagged data for parameters such as directionality, movement amplitude, and / or response duration may be extracted for analysis. The extracted features are further processed by the trained classification model for peripheral or central vestibular dysfunction, such as vertigo or imbalance.

[0163] In at least one embodiment, when the trained classification model is configured for dementia detection, the input layer 304 may be configured to extract tagged data associated to the subject’s values for reduced look-to-stare ratios, impaired smooth pursuit, and / or slower or less frequent saccades. Tagged data associated to pursuit gain, interruptions, and saccadic dynamics may also be extracted. The tagged data associated to reduced reflexive responses to optokinetic stimuli may also be extracted for processing by the classification model for neurodegenerative patterns.

[0164] As another example, in at least one embodiment, when the trained classification model is configured for concussion or mild traumatic brain injury (mTBI) detection, the input layer 304 may extract tagged data related to delayed or erratic saccades, gaze instability, and / or irregular reflexes. In at least one embodiment, the tagged data associated to the saccadic speed, reset delay, and interruptions during smooth pursuit can also be extracted. The extracted feature data is processed by the trained classification model to identify disruptions in reflexive and voluntary eye movements to detect traumatic brain injuries.

[0165] As another example, in at least one embodiment, when the trained classification model is configured to detect fatigue or intoxication, the input layer 304 may extract subject’s tagged data associated to reduced pursuit gain, gaze-evoked nystagmus,and / or saccadic intrusions. The tagged data associated to reflexive responses to optokinetic stimuli is extracted for further processing to detect reversible impairments.

[0166] . The input layer 304 further processes the classified data for feature extraction to generate feature data 306. The classified data may include multiple classified data values 306, corresponding to a specific tag or classification such as, but not limited to pursuit gain, nystagmus frequency, or saccade speed, for example.

[0167] The inference layer 308 receives and processes the feature data 306 received from the input layer 304 by applying a trained neural network to analyze the patterns associated with a specific condition. The feature data values 306 represent normalized and tagged classifications, such as pursuit gain, interruption frequency, and / or nystagmus amplitude, for example, that are relevant to the condition being assessed by the trained classification model. The inference layer 308 can implement a network of interconnected nodes, where the node(s) apply weights to the extracted feature data values 306 based on the significance of the features for the classification task. The weights can be assigned during the training phase of the classification model using labeled datasets.

[0168] The node(s) in the inference layer 308 performs a computational operation to process the feature data 306. The node(s) combine(s) the incoming weighted feature data and applies an activation function to introduce non-linearity. The inference layer 308 is configured to process complex relationships between the feature data values 306 and the associated conditions. For example, if the feature data 306 include high pursuit gain and frequent interruptions, the node activation may reflect a likelihood of neurodegenerative conditions such as dementia. The activation functions can include sigmoid, ReLU (rectified linear unit), or other functions suitable for the classification model’s architecture.

[0169] The inference layer 308 may propagate the processed outputs from the node(s) to other layer(s) within the neural network of the inference layer 308. For example, the inference layer 308 may identify that a combination of asymmetric nystagmus and delayed saccades correlates with stroke-related conditions. The aggregation process provides for holistic processing of the feature data 308.

[0170] Weights assigned to the feature data 308 can be dynamically adjusted during the training phase to improve the trained classification model’s 300 performance. The weight(s) represent the relative importance of a feature data for the classification task.For instance, the trained model 300 may assign a higher weight to feature data associated to interruption frequency when analyzing smooth pursuit for dementia detection. Similarly, the feature data associated to nystagmus directionality may receive higher weight in classification model configured for vestibular disorder analysis.

[0171] In at least one embodiment, the inference layer 308 processes the corresponding stimulus data included in the feature data. The stimulus data provides context by including the characteristics of the visual stimuli that elicited the recorded eye movements. Parameters such as stimulus type (e.g., vertical or horizontal stripes), speed, color, direction, and display conditions can be integrated into the feature data. The inference layer 308 processes the association between the stimulus characteristics and the tagged eye movement data (feature data) to identify patterns that correspond to specific conditions. For example, the inference layer 308 may be configured to analyze whether a high-frequency nystagmus response aligns with fast-moving horizontal stimuli. The processing of stimulus data includes applying weights to the stimulus parameter(s) based on its contribution to the classification task. The inference layer 308 nodes can also be configured to map the eye movement responses to the corresponding stimulus properties. For example, the inference layer 308 may compare the velocity of a subject’s smooth pursuit movements with the velocity of the moving stimulus to evaluate pursuit gain.

[0172] The inference layer 308 generates evaluation data after processing the feature data through multiple nodes. The evaluation data indicates the extent to which the feature data matches the patterns associated with the condition being examined. The evaluation data may include a classification score or probability, representing the likelihood of the condition’s presence. For instance, the inference layer may produce a classification score of 0.85, indicating an 85% probability that the subject exhibits features consistent with dementia.

[0173] The probability level in the evaluation data generated by the inference layer 308 can be based on patterns provided during the training phase. Labeled datasets used for training include feature tags such as pursuit gain, interruption frequency, and / or saccadic delay, mapped to known conditions like cognitive decline, vestibular dysfunction, or fatigue. The inference layer 308 can finetune based on the trained patterns to evaluate the incoming feature data and assign confidence levels to the classification results.

[0174] In at least one embodiment, the inference layer 308 generates inferential data that identifies additional eye movement features or tags potentially correlated with the condition being processed by the trained classification model. In such embodiments, based on patterns in the training data and the subject’s feature data measurements, the inference layer 308 can detect new associations between previously untagged parameters and the examined condition. The inference layer 308 generates inferential data by analyzing the weights and activations of nodes involved in the feature evaluation. In at least one embodiment, correlation matrices and / or attention mechanisms may be implemented within the neural network 308 that may identify features that exhibit statistical or probabilistic relevance to the condition. In at least one embodiment, the inferential data generated by the inference layer 308 configures the input layer 304 to dynamically adjust the tagging and extraction process. For instance, the input layer 304 may be trained to prioritize and extract new tags, such as eyelid dynamics or asymmetry in nystagmus, based on the inferential data. The inferential data improves the trained classification model 300 to refine its analysis in real time and identify new correlations.

[0175] The trained classification model includes an output layer 310 configured to receive evaluation data from the inference layer 308. The output layer 310 is configured to generate assessment data for the classification model 300. The evaluation data received from the inference layer 308 may include a classification score or probability and intermediate outputs that represent the aggregated contributions of the analyzed feature data. The output layer 310 processes the evaluation data to generate interpretable results in the form of assessment data. The assessment data may be generated based on a mapping of analyzed feature patterns derived from the feature data provided by the input layer 304 to specific classification categories relevant to the condition being examined by the model. In at least one embodiment, the inference layer 308 generates inferential data identifying additional eye movement features or tags correlated with the condition based on node weights, activations, and / or correlation patterns during inference. The inferential data includes a ranked set of candidate condition-specific tags and associated relevance values. The multi-model architecture 106 stores the candidate tags as updated relevancy parameters for the corresponding trained classification model. During a subsequent execution, the input layer 304 applies the updated relevancy parameters to adjust thetagging and feature extraction process to include the candidate tags in the classified data and feature data 306. Thus, the tag set is adaptively refined over time based on inferential outputs without requiring manual retagging of prior datasets.

[0176] In at least one embodiment, the output layer 310 may apply decision thresholds to the classification score received from the inference layer 308. The decision thresholds determine the classification outcome based on the probability scores. For example, a classification score exceeding a predefined threshold, such as 0.85, may correspond to a positive detection of the condition being analyzed. Alternatively, a score below the threshold may indicate the absence of the condition. The decision thresholds can be dynamically defined during the training phase of the classification model or configured based on user input.

[0177] The assessment data may include the classification category, associated confidence metrics, and other relevant condition-related insights. For example, in a classification model 300 configured for dementia-related analysis, the output layer 310 may generate an assessment data indicating “Condition Detected” with a confidence score of 85%, accompanied by metadata and feature data values such as a normalized pursuit gain value of 0.75 and an interruption frequency of 3 events per second. Additional annotations may provide specific eye movement patterns that contributed most to the classification result. Additionally, the temporal association between the stimuli data and the corresponding eye movement data can be included in the assessment data.

[0178] In at least one embodiment, the output layer 310 may also be configured to format the assessment data for integration with external systems or display devices. For example, the assessment data may be encoded in standardized formats, such as JSON, XML for interoperability with decision-support tools, reporting systems, or cloud-based platforms. For instance, the output layer 310 may transmit the formatted results to an administrator’s workstation, a clinician’s tablet, or a remote server for visualization or further analysis. The formatted results can include condition-related summaries or visual representations of feature trends, enabling actionable insights.

[0179] In at least one embodiment, the assessment data generated by the output layer 310 may include pattern correlation data. The pattern correlation data may improve the assessment features of the classification model. The pattern correlation may includelongitudinal tracking or a graphical representation of eye movement parameters, comparative analysis against baseline measurements, or identification of overlapping abnormalities. The output layer may track changes in the feature data over time to generate longitudinal tracking data to provide a representation of the progression or improvement of the condition. For example, in a vestibular disorder classification model, the output layer 310 may correlate patterns of off-axis nystagmus and abnormal optokinetic after-nystagmus (OKAN) to identify the coexistence of peripheral and central vestibular dysfunctions. The pattern correlation data may include annotations to indicate the correlational findings and suggest potential follow-up interventions. The pattern correlation data may provide predictive analytics generated by regression models or trend analysis techniques. The output layer 310 can be configured or trained to analyze gradual changes in pursuit gain, interruption frequency, and / or saccadic dynamics over time to predict the progression of conditions such as dementia, Parkinson’s disease, or vestibular decline.

[0180] In at least one embodiment, the pattern correlation data can be presented in the form of visualizations, such as trend graphs, heat maps, or comparative tables. For example, longitudinal changes in nystagmus amplitude can be visualized as a line graph. The comparative baseline analysis can also be visualized in tabular form with highlighted deviations.

[0181] In at least one embodiment, the output layer 310 is configured to generate administration data that provides actionable guidance based on the condition detected by the trained classification model. The administration data may include treatment recommendations, next steps, and / or referrals based on the evaluation data. For example, upon determining a positive classification, the output layer 310 may reference a pre-defined database or decision-support system (not shown) that is accessible by the trained classification model. The database may be configured to include evidence-based guidelines, condition-related protocols, and / or treatment algorithms associated for the detected condition. For example, in a dementia-related trained classification model, the administration data may recommend interventions such as cognitive rehabilitation exercises, specific pharmacological treatments, and / or referrals to a neurologist. Theadministration data can be derived from the classification score and severity level for personalized condition-specific guidance.

[0182] Additionally, in at least one embodiment, the administration data may include condition-specific annotations describing the rationale for the recommended actions. The annotations may include highlighting informative contributing features from the evaluation data. For example, the administration data may indicate that frequent interruptions in smooth pursuit and reduced pursuit gain contributed to dementia and recommend targeted cognitive therapies.

[0183] In at least one embodiment, the system 300 is implemented so that it uses a single classification model designed to process all captured eye movement data without the need for selecting or switching between multiple models. The input layer operates as the initial processing stage, receiving the raw eye movement data collected during optokinetic reflex (OKR) tests or similar visual stimuli exposures. The input layer assigns tags to the data point(s) based on predefined classifications. The tagged classifications include measurable parameters such as any combination of amplitude, frequency, velocity, symmetry, and directionality of eye movements, and optionally auxiliary features like gaze position and / or eyelid dynamics.

[0184] The input layer further converts the tagged classifications into normalized numerical values to generate classified data. The normalization process adjusts the raw values to fall within a standardized range configured for the classification model. The tagged classifications and corresponding normalized values are combined to create a comprehensive dataset.

[0185] Some or all of the comprehensive dataset is then forwarded to the inference layer for processing. The inference layer is implemented as a neural network trained on labeled datasets that associate tagged features with specific conditions. The trained neural network comprises a series of interconnected nodes, where the node(s) apply weights to the incoming data based on the significance of the feature datum to the condition(s) assessed by the classification model. The weights can reflect the relative importance of different eye movement features in identifying the conditions. For instance, higher weights may be assigned to parameters such as pursuit gain or interruption frequency when analyzing smooth pursuit abnormalities associated with neurodegenerative disorders.

[0186] The inference layer can perform a series of computational operations to evaluate the tagged classifications. The computational operations include weighted summation, where the incoming feature data is combined based on assigned weights, followed by the application of activation functions. Activation functions, such as sigmoid or rectified linear unit (ReLU) can be implemented to introduce non-linearity to enable the network to identify complex patterns within the dataset. For example, if the tagged data includes high-frequency nystagmus and asymmetric movement patterns, the activation functions may highlight the patterns as indicative of vestibular dysfunction or stroke-related conditions.

[0187] The inference layer propagates the processed outputs through multiple connections of nodes within the neural network to aggregate the contributions of all tagged features. The multi-node processing allows the system to analyze the interactions between different parameters and identify holistic patterns associated with specific condition(s).

[0188] Based on the aggregated analysis, the inference layer generates a probability score that represents the likelihood of a condition being present. The probability score is calculated by evaluating how closely the processed feature data matches patterns learned during the training phase of the neural network. The inference layer can compare the probability score to a predefined decision threshold to determine whether the condition is detected. For example, a score exceeding 0.85 may indicate an 85% likelihood of a condition such as dementia, while a score below the threshold may signify the absence of the condition.

[0189] In at least one embodiment, the inference layer generates a plurality of probability scores for a predefined list of potential conditions and determines the condition associated with the highest probability score that meets or exceeds the decision threshold. For instance, if the model generates probability scores of 0.85 for dementia, 0.75 for vestibular dysfunction, and 0.65 for fatigue, the model selects dementia as the detected condition because it has the highest probability score exceeding the threshold. The selected condition, along with its probability score and contributing feature insights, is then transmitted to the output layer for further processing and report generation.

[0190] In the various embodiments described herein, the trained classification models can be implemented using certain artificial intelligence (Al) technologies to improve the accuracy. For example, one or more of the trained classification models used herein can be based on deep learning models, such as convolutional neural networks (CNNs). The input layer can utilize CNNs to tag, normalize, and segment the eye movement data based on pre-trained features. The inference layer applies probabilistic analysis through multi-layered CNN architectures to detect subtle patterns indicative of specific neurological conditions, such as optokinetic dysreflexia or vestibular dysfunction.

[0191] In at least one embodiment, one or more of the trained classification models can be implemented using recurrent neural networks (RNNs) or long short-term memory (LSTM) networks. In such cases, the inference layer can implement LSTMs to evaluate the temporal dependencies within the feature data to enable the detection of anomalies such as delayed saccadic resets or irregular pursuit gains.

[0192] In at least one embodiment, in addition to neural networks, one or more of the trained classification models can be implemented using a gradient boosting machines (e.g., XGBoost) or random forests to improve decision-making in the output layer. The output layer can integrate the probabilistic outputs from various inference pathways, weigh the contributions of the feature(s), and generate actionable insight. Furthermore, explainable Al (XAI) techniques can be provided at the output layer to provide user with an understanding of how specific eye movement parameters influenced the classification decision.

[0193] Reference is now made to FIG. 4, which shows a flowchart 400 of a method for feature extraction and assessment, according to an example embodiment. The method 400 can be implemented by the multi-model architecture of FIG. 1.

[0194] At 402, the method 400 includes generating, a plurality of visual stimuli configured to elicit a plurality of eye movement data in a subject. The plurality of visual stimuli can be generated to comprise any combination of one or more patterns, one or more speeds, one or more directions, and one or more colors designed to trigger specific reflexive eye movements, such as optokinetic reflexes or smooth pursuit.

[0195] For example, the visual stimuli including alternating striped patterns (e.g., black and white stripes), moving dot arrays, and / or shapes are designed to elicit nystagmus.The stimuli can be configured with variable widths, speeds, and / or directions, such as horizontal, vertical, and / or diagonal movements. In some embodiments, color variations, such as alternating red and blue stripes, may be presented to assess chromatic sensitivities or reflex variations. The stimuli can also include dynamic elements, such as accelerating or decelerating movement patterns.

[0196] The display device presenting the stimuli may include display platforms depending on the testing context. Portable devices, such as tablets, smartphones, or laptops comprising display screens or high-resolution screens, may be used for remote or point-of-care applications. Specialized devices, such as virtual reality (VR) headsets or visors, can be used to provide immersive full-field stimulation to ensure accurate elicitation of reflexive movements. The devices may integrate or be connected to cameras or infrared sensors for capturing and real-time tracking of eye movements.

[0197] At 404, the method 400 includes receiving, the plurality of eye movement data comprising optokinetic reflex (OKR) responses. The plurality of eye movement data is captured during presentation of the plurality of visual stimuli to the subject.

[0198] The eye movement data may be determined from images of the subject’s eyes acquired during presentation of the visual stimuli by obtaining measurements based on those images to obtain values for parameters such as any combination of amplitude, frequency, velocity, pursuit gain, interruptions, symmetry, and directionality of eye movements. Additionally, in at least one embodiment, the method 400 may include determining other parameter values from auxiliary data associated with the eye movement data, such as gaze position and / or eyelid dynamics. The received data can represent the subject’s reflexive or voluntary eye movements elicited during the visual stimulus presentation.

[0199] At 406, the method 400 may include selecting a trained classification model by performing a probabilistic analysis on a set of selection parameters applied to the plurality of eye movement data. The set of selection parameters may comprise characteristics of the plurality of eye movement data. Alternatively, in at least one embodiment, one trained classification model may instead be used as described earlier.

[0200] The selection parameters represent measurable attributes and contextual information for determining the most appropriate trained classification model to processthe input data. The selection parameters may comprise characteristics derived from the eye movement data, including specific movement types (e.g., a combination of saccadic movements, smooth pursuit, and / or reflexive gaze stabilization) and / or tagged classifications (e.g., pursuit gain, and / or interruption frequency). The selection parameters can provide an indication of potential conditions. For instance, high interruption frequency or reduced pursuit gain may suggest cognitive decline or neurological dysfunction. Additionally, characteristics of the stimulus data, such as its speed, direction, and / or pattern configuration, may also be included as selection parameters to evaluate the assessment process assessed for the given data. Further, user-provided inputs, including suspected conditions or prior medical history, may perform as additional selection parameters. The selection parameters can be received as user inputs. Alternatively, selection parameters can be determined during the training phase of the trained selection model by processing labeled datasets of eye movement patterns and their associated conditions. During training, the trained selection model is obtained from evaluating correlations between specific features (e.g., pursuit gain, saccadic delay) and known conditions to identify the parameters most relevant for model selection during actual use.

[0201] The method 400 includes performing a probabilistic analysis by applying the set of selection parameters to the plurality of eye movement data. The probabilistic analysis is implemented by the trained selection model, which may be artificial intelligence-based and trained on labeled datasets of eye movement patterns and associated outcomes. The method 400 includes assigning weights to the selection parameter(s) based on its significance to the respective condition(s). For example, a tagged feature or a selection parameter such as reduced pursuit gain may receive a higher weight in the context of dementia-related classification models compared to vestibular dysfunction models. The method 400 may include calculating a likelihood score for one or more of the trained classification model(s) in the multi-model architecture by aggregating the weighted contributions of the parameters.

[0202] For instance, the probabilistic analysis may determine a high likelihood score for a trained classification model configured for dementia assessment if the eye movement data indicates frequent interruptions in smooth pursuit and the stimulus data correspondsto high-speed horizontal stripes. The method generates a selection signal that identifies the trained classification model with the highest likelihood score based on the probabilistic analysis. The selection signal is transmitted to the multi-model architecture to activate the corresponding classification model.

[0203] At 408, the method 400 includes segregating the plurality of eye movement data to generate tagged classifications by assigning predefined tags to measurable parameters of the eye movement data and converting the tagged classifications into normalized numerical values to generate classified data.

[0204] The categorization or segregation process includes analyzing the captured eye movement data to identify measurable parameters such as amplitude, frequency, velocity, symmetry, and / or directionality. The parameter(s) are assigned a predefined tag that categorizes the parameters into distinct classifications, including nystagmus, saccades, smooth pursuit, look-to-stare ratios, and / or palpebral fissure area (PFA). For instance, nystagmus-related parameters such as amplitude and frequency are tagged to facilitate the identification of reflexive patterns. Similarly, the saccades are tagged with parameters such as speed, accuracy, and / or reset delay. The tagging process provides that the data point(s) in the eye movement dataset is systematically labeled to enable structured analysis.

[0205] The method further includes assigning tags to classify abnormalities or irregularities in reflexive or voluntary eye movements. For example, optokinetic dysreflexia (OKD), characterized by asymmetric or absent reflexive eye movements, may be tagged as an abnormal classification indicative of neurological or physiological impairments. Similarly, parameters related to saccadic movements, such as delayed reset or erratic trajectories, are tagged to indicate potential neurodegenerative disorders or traumatic brain injuries.

[0206] Once the tagged classifications are assigned, the method 400 includes converting the tagged data into normalized numerical values to generate classified data. The normalization process includes scaling continuous measurements to a standardized numerical range for compatibility with the classification model’s processing requirements. For example, if a pursuit gain value is measured as 0.75, the pursuit gain tag may benormalized to a scale of 0 to 1, while interruptions in smooth pursuit are encoded as discrete event frequencies (e.g., three interruptions per second).

[0207] In at least one embodiment, the method 400 may also include associating the normalized values with metadata describing the stimuli that elicited the corresponding eye movements. For example, the classified data may include entries such as “Tag: Pursuit Gain, Value: 0.75, Stimulus: Horizontal Stripes, Timestamp: T1.” The association provides temporal and spatial context to the classified data to provide a complete representation of both the subject’s responses and the triggering conditions.

[0208] Additionally, the tagging process may include assigning symbolic or numerical identifiers to the classified data for efficient organization and retrieval. For example, pursuit gain may be assigned a tag identifier such as “1,” and saccadic speed may be tagged as “2.” The identifiers group the normalized data values under their respective classifications.

[0209] At 410, the method 400 includes generating feature data by extracting tagged classifications from the classified data based on relevancy parameters associated with the trained classification model, the relevancy parameters comprising condition-specific tags.

[0210] The relevancy parameters refer to the condition-specific tags that correspond to measurable eye movement characteristics for assessing the condition assessed by the classification model. The feature extraction process isolates the tagged classifications based on the relevancy parameters, filtering out non-essential or irrelevant data points in the classification data.

[0211] The relevancy parameters may be determined during the training phase of the trained classification model. The relevancy parameters can be dynamically assessed during the training based on the labeled datasets correlating specific eye movement features with particular conditions. For example, for a trained classification model configured to detect dementia, the relevancy parameters may include tags such as pursuit gain, interruption frequency, and saccadic reset delay, which are associated with neurodegenerative patterns. The input layer applies the relevancy parameters to extract the tagged classifications that align with the assessment focus of the model.

[0212] The method 400 may include dynamically refining the relevancy parameters through adaptive learning algorithms in the classification model. The adaptive learning process uses historical data and outcomes from prior analyses to identify correlations between tagged features and the targeted condition. For instance, if a previously untagged feature, such as eyelid asymmetry, is found to have condition-related value for a vestibular disorder, the relevancy parameters are updated to include the previously untagged feature.

[0213] In at least one embodiment, the relevancy parameters are based on user inputs or clinician-defined criteria to determine the selection of relevancy parameters. A clinician may prioritize tags based on a subject’s medical history or symptoms. For example, in cases where a subject exhibits symptoms indicative of stroke, the clinician may emphasize tags such as asymmetric nystagmus or delayed reflexes.

[0214] By applying the relevancy parameters, the method 400 reduces the computational burden on subsequent layers, such as the inference layer. Extracting the relevant tagged classifications may reduce the volume of data processed, improving efficiency and enabling faster, more focused analysis. For example, when the trained classification model is configured for nystagmus detection, the method 400 isolates tags related to amplitude, frequency, direction, and / or symmetry, while excluding unrelated features such as palpebral fissure measurements.

[0215] The feature data generated by the method 400 includes structured representations of the extracted tagged classifications and their normalized values.

[0216] At 412, the method 400 includes determining a condition when a probability score meets a decision threshold. The probability score is generated by the trained neural network configured to process the tagged classifications in the feature data with assigned weights to identify patterns indicative of the condition.

[0217] The method 400 includes applying the tagged classifications from the feature data to a network of interconnected nodes in the trained neural network. The node(s) combine the tagged classifications with their assigned weights. The weight represents the significance of the tagged classification in identifying the specific condition. The method 400 further includes processing the combined weighted values at the nodes throughactivation functions, such as sigmoid or rectified linear unit (ReLU) to generate node outputs that reflect the likelihood of specific condition patterns.

[0218] The method 400 includes propagating the node outputs through additional interconnections of nodes of the trained neural network to evaluate complex relationships between the tagged classifications and the condition being analyzed. The method 400 includes aggregating the processed outputs from multiple nodes to calculate the probability score, which indicates the extent to which the feature data matches the patterns associated with the condition. For instance, the probability score may be calculated by analyzing patterns such as high pursuit gain and frequent interruptions, which are indicative of neurodegenerative disorders such as dementia.

[0219] The method 400 includes comparing the calculated probability score to a predefined decision threshold to determine the presence of the condition. For example, a probability score exceeding a threshold of 0.85 may indicate a positive indication of the condition. The threshold values may be predefined during the training phase of the neural network or dynamically configured based on input from users or clinicians.

[0220] The weights assigned to the tagged classifications in the feature data can be dynamically adjusted during the training phase. The weight(s) can represent the relative importance of the tagged classifications in the feature data for the classification task. For example, the trained model 300 may assign a higher weight to feature data associated to interruption frequency when analyzing smooth pursuit for dementia detection. Similarly, the feature data associated to nystagmus directionality may receive higher weight in classification model configured for vestibular disorder analysis.

[0221] In at least one embodiment, the probability score is generated by processing the corresponding stimulus data included in the feature data. The stimulus data provides context by including the characteristics of the visual stimuli that elicited the recorded eye movements. Parameters such as stimulus type (e.g., vertical or horizontal stripes), speed, color, direction, and / or display conditions can be integrated into the feature data. The neural network processes the association between the stimulus characteristics and the tagged eye movement data to identify patterns that correspond to specific condition(s).

[0222] In at least one embodiment, the method 400 includes generating inferential data that identifies additional eye movement features or tags potentially correlated with thecondition being processed by the trained classification model. The method 400 further comprises generating inferential data identifying a plurality of additional condition-specific tags based on the analysis of node weights and correlation patterns by the neural network. The inferential data provides probabilistic relevance of untagged parameters to the condition. Based on patterns in the training data and the subject’s feature data measurements, the trained neural network can detect new associations between previously untagged parameters and the examined condition. The trained neural network generates inferential data by analyzing the weights and activations of nodes involved in the feature evaluation. Correlation matrices and attention mechanisms within the trained neural network may identify features that exhibit statistical or probabilistic relevance to the condition, even if they were not part of the initial feature extraction process.

[0223] At 414, depending on the embodiment, the method 400 may optionally include generating administration data for the detected condition based on a pre-defined database configured with treatment protocols specific to the condition. The administration data may comprise treatment recommendations.

[0224] The administration data provides guidance based on the condition detected by the neural network. The administration data may include treatment recommendations, next steps, or referrals based on the evaluation data. Upon determining a positive classification, the trained neural network references a pre-defined database or decisionsupport system (not shown) that is accessible to the trained classification model. The database is trained with evidence-based guidelines, condition-related protocols, and / or treatment algorithms associated with the detected condition.

[0225] The method 400 includes generating assessment data based on a classification score or probability and intermediate outputs that represent the aggregated contributions of the analyzed feature data. The assessment data maps analyzed feature patterns derived from the feature data provided by the neural network to specific classification categories relevant to the condition being examined by the model. The decision thresholds determine the classification outcome based on the probability scores. The assessment data includes the classification category, optionally associated confidence metrics, and optionally other relevant condition-related insights. For example, in a neural network configured for dementia-related analysis may generate an assessment dataindicating “Condition Detected” with a confidence score of 85%, accompanied by metadata and feature data values such as a normalized pursuit gain value of 0.75 and an interruption frequency of 3 events per second. Additional annotations may highlight specific eye movement patterns that contributed most to the classification result. Additionally, the temporal association between the stimuli data and the corresponding eye movement data can be included in the assessment data.

[0226] In at least one embodiment, the assessment data generated may also include pattern correlation data. The pattern correlation data may improve the assessment features of the trained classification model. The pattern correlation may include longitudinal tracking of eye movement parameters, comparative analysis against baseline measurements, and / or identification of overlapping abnormalities. The pattern correlation data can be presented in the form of visualizations, such as trend graphs, heat maps, or comparative tables.

[0227] Reference is now made to FIG. 5, which illustrates a block diagram of an example user apparatus 500 for determining eye movement data. The user apparatus can 500 can be the user apparatus 102 of Fig. 1. The apparatus 500 is configured to capture and process reflexive eye movement data elicited during optokinetic reflex (OKR). The apparatus 500 includes a display device 502, an imaging device 504, a stimulus generator for generating a visual stimulus 506, and a visor 508.

[0228] The display device 502 is configured to present optokinetic stimuli configured to elicit reflexive eye movements. The visual stimuli may include dynamic, high-contrast patterns, such as rectangular bands or stripes, moving in predefined directions across the display. The patterns are configured to stimulate smooth pursuit eye movements (SPEM) followed by saccadic resets. The stimuli are designed for producing the characteristic optokinetic nystagmus waveform. The display device can include configurable settings, to provide for adjustments to stimulus speed, color, and directionality. For example, horizontal stimuli moving left to right can assess horizontal gaze stability, while vertical stimuli are used to evaluate vertical reflex mechanisms. The display device 502 provides for consistent and uniform presentation of stimuli.

[0229] The imaging device 504 can be a video camera configured to capture real-time video of the subject’s eye movements in response to the displayed stimuli. In at least oneembodiment, a processor connected to the imaging device 504 is configured to provide waveform tracking and analysis from the raw video captured from the imaging device. The processor receives and processes the captured video data and may extract eye movement data, including tracking waveforms such as pupil displacement over time, nystagmus amplitude, and palpebral fissure height. The camera is positioned to provide precise tracking of pupil motion, capturing image data from which various measurements can be made such as gaze direction, saccadic resets, and / or deviations from resting positions. In an embodiment, the eye movement data to be obtained from the image data is determined by the condition being assessed by the trained classification model. The imaging device 504 may integrate video processing capabilities to detect and isolate eye features. The imaging device 504 can transmit the raw video or extracted waveform data to a processing device to perform various measurements, including generating featurerich outputs, which may all be saved as eye movement data. The processing device can communicate with the multi-model architecture 106.

[0230] The stimulus generator, communicatively connected to the display device 502, can be a processing device (not shown) such as a processor. The stimuli are calibrated to produce reflexive eye movements consistent with the optokinetic reflex mechanism. The periodic motion of the patterns can stimulate cortical, brainstem, and cerebellar circuits used in vision and oculomotor control. The imaging device 504 supports the use of full-field stimuli to capture the subject’s visual field to elicit reflexive responses. Additionally, targeted stimuli can be presented to focus on specific reflex pathways. For example, high-speed horizontal stimuli can be used to determine optokinetic dysreflexia, a condition characterized by impaired gaze stabilization.

[0231] In at least one embodiment, the apparatus 500 may instead use a visor 508, rather than the display 502, where the visor displays the visual stimuli and has one or more cameras for obtaining images of the subject’s eyes during testing. The use of a visor 508 may improve testing accuracy by aligning the subject’s visual field and minimizing interference from ambient light. The visor 508 standardizes the subject’s field of view, focusing their attention on the stimuli presented on the display device while reducing extraneous visual distractions. For tests conducted in darkness, such as optokinetic afternystagmus (OKAN) assessments, the visor 508 is configured to block external light. Byisolating the subject’s visual environment, the visor improves the reliability and reproducibility of the test results.

[0232] In at least one embodiment, the apparatus 500 may be implemented using a smartphone or tablet device. The smartphone’s display can operate as the stimulus presentation medium. The smartphone’s integrated camera may capture video data of the subject’s eye movements. The optokinetic reflex test may include moving contrasting rectangular bands on the display to elicit reflexive eye movements. The smartphone processes pupil motion relative to resting positions and evaluates the responses against reference waveforms stored in its memory. The processing device or application on the smartphone may trigger alerts when deficient reflex responses are detected, such as the absence of measurable nystagmus or waveforms deviating from normal patterns. The smartphone may be further connected to the multi-model architecture for advanced further data processing and classification.

[0233] In at least one embodiment, the user apparatus 500 may be configured to capture and process eye movement data during an optokinetic reflex (OKR) test to determine the presence or strength of the reflex. The user apparatus 500 is configured to display optokinetic stimuli on the display device 502 which may include rolling striations with a motion range of 3-4° to elicit and measure reflexive eye movements indicative of alertness. The stimuli are presented in such a way that the visual response, characterized by nystagmus, can be measured to assess the subject’s alertness level. An unexpected lack of nystagmus, or attenuated reflex, can be considered as an indication for reduced alertness.

[0234] The imaging device 504, such as a video camera, captures real-time input video signals of the subject’s eye movements during the test. The imaging device 504 can be positioned to track pupil motion with high accuracy. The captured signals are transmitted to a processing device for further analysis. The processing device can determine whether the eyes of the subject are present in the captured frame and analyze the optokinetic nystagmus waveform.

[0235] In at least one embodiment, the imaging device 504 captures images which may be processed, e.g., by a processor, to measure the vertical distance between the upper and lower eyelids to capture changes in the palpebral fissure area (PFA) over time duringvisual stimuli. The imaging device 504 can be focused on the subject’s eyelid region to continuously record eyelid movements and subtle variations in eyelid separation. The captured image data is obtained during optokinetic reflex (OKR) testing, where eyelid dynamics act as a reliable proxy for reflexive and voluntary eye movements. The measured PFA data provides an advantage over traditional pupil tracking by delivering insights to neurological and physiological conditions that impact eyelid behavior. For example, the measured PFT data may indicate PFA abnormalities such as asymmetry, delayed reflexive responses, or lid lag, enabling the detection of conditions like hyperthyroidism, cranial nerve palsy, stroke, or fatigue. Accordingly, the visual testing and captured image data includes reflexive eyelid responses that are subtle markers of fatigue, intoxication, or concussion, providing broader neurological insights. Unlike pupil tracking, which relies on controlled lighting environments, PFA measurements can be less affected by lighting conditions and the images captured for obtaining the PFA measurements can be done using standard video cameras or infrared sensors. Additionally, the apparatus 500 along with the imaging device 504 provides for accessible and less invasive PFA tracking, avoiding the need for specialized or intrusive equipment.

[0236] In at least one embodiment, the apparatus 500 measures the Look-to-Stare Ratio by monitoring palpebral area changes and eye movements during vertical optokinetic stimulation (OKS). The imaging device 504 captures real-time video of the subject’s eyelid movements as they respond to controlled visual stimuli displayed on a screen. The visual stimuli, which may include moving vertical stripe patterns, elicit alternating smooth pursuit and saccadic reset movements, collectively known as optokinetic nystagmus (OKN). The apparatus 500 can process the image data to track subtle variations in the palpebral aperture, or the space between the upper and lower eyelids, over time as the subject’s eyes move in response to the visual stimulus. The imaging device 504 can be positioned for alignment with the subject’s visual field for tracking of eyelid position changes. A video feed is captured, to continuously record frames of the eyelid movements. The apparatus 500 can focus on identifying and recording minute adjustments in eyelid separation, correlating these changes with the underlying eye movement patterns. The imaging process captures smooth pursuit movements, which may be characterized by steadiereyelid dynamics, as well as saccadic resets, marked by sharper, discrete adjustments in the palpebral aperture.

[0237] The raw video data captured by the imaging device 504 can be transmitted to a connected processing device. The processing device may preprocess the data to extract parameters, such as eyelid position, movement frequency, and amplitude. The processing device can identify the transitions between “look” and “stare” responses by analyzing the captured waveforms generated from the eyelid movement data.

[0238] The processing device is configured to detecting and classifying various types of nystagmus waveforms, including pendular responses (pure, asymmetric, and with foveating saccades), unidirectional jerk responses (pure, extended foveating, pseudo cycloid, and pseudo jerk), and bidirectional pseudo pendular responses (triangular, bidirectional jerk, and others). Dual jerk responses can also be identified as part of the advanced classification features. By extracting a specific optokinetic waveform from the captured data, the processing device can process whether the measured responses fall within a range indicative of normal alertness or are associated with abnormalities.

[0239] The stimulus generator is configured to present the stimuli within the field of view aligned to the subject’s gaze direction. The processing device may determine the optokinetic nystagmus waveform from the pupil response captured in the input video signal (which contains a series of video images or image data).

[0240] The processing device can perform several computational tasks on the input video signal (image data) received from the imaging device 504 to obtain eye movement data. For example, the processing device can calculate eye movement velocities by applying derivative calculations to displacement data obtained from the image data. Further, the processing device can categorize the eye movements into slow-phase and quick-phase components based on predefined velocity thresholds. Slow-phase movements are identified when the velocity remains within a threshold range for a specified duration, while quick-phase movements are detected when the velocity exceeds the threshold. Linear regression techniques can be employed to fit a regression line to the data of the slow-phase movement. The processing device can generate a series of linear regression lines for the identified slow-phase movements, enabling precise characterization of the optokinetic nystagmus waveform.

[0241] The integration of derivative calculations, threshold adjustments, and regression analysis allows the processing device to adaptively refine its preliminary analysis. By dynamically selecting threshold ranges that minimize deviations between regression lines and the corresponding eye movement data, the processing device can provide high accuracy in detecting and categorizing eye movement patterns. These classifications may then be included in the eye movement data that is processed by one or more trained classification models described herein.

[0242] In at least one embodiment, the processing device may be configured to perform pre-processing of the image data received from the imaging device 504 before preparing the eye movement data for analysis by the model selection module and / or one of the trained classification models. The raw video signals captured by the imaging device 504 may be analyzed to extract features such as pupil position, gaze direction, and movement velocity. The processing device can apply video processing algorithms to isolate the pupil’s trajectory and remove noise or irrelevant visual artifacts that may interfere with accurate analysis. The pre-processing provides that the extracted data reflects the subject’s reflexive responses to the stimuli displayed on the display device 502.

[0243] In at least one embodiment, the processing device can segment the eye movement data into discrete intervals corresponding to specific periods of stimulus presentation. For instance, the segment(s) can be labeled with the characteristics of the associated visual stimulus, such as movement direction, speed, or contrast pattern. The segmentation provides that the eye movement data is temporally aligned with the stimuli, providing that the model selection module processes both the reflexive responses and their contextual triggers during analysis. The structured representation of eye movement data, tagged with relevant stimulus metadata, is primed for the model selection module.

[0244] Reference is now made to FIG. 6, which illustrates an example of optokinetic stimulation and eye movement data capture 600. The stimulation may be processed by the user apparatus 500 of FIG. 5.

[0245] The visor provides an immersive full-field optokinetic stimulation. The stimulus comprises of repetitive high-contrast vertical patterns presented to the user’s visual field to elicit vertical optokinetic reflex (OKR) responses. The optokinetic stimuli are configuredto activate the neuro-oculomotor pathways responsible for gaze stabilization and reflexive eyelid movements.

[0246] As shown in FIG. 6, the imaging device 504 is integrated into the visor films the subject’s eye movements and dynamically tracks the palpebral fissure height (PFH) during the duration of the optokinetic stimulation. The palpebral fissure height may refer to the vertical distance between the upper and lower eyelids, measured in response to a visual stimulus. The processing device may process the input video data to extract eye movement waveforms, such as PFH versus time (T), and generates detailed plots that represent the reflexive responses.

[0247] The processing device may be configured to detect deviations from expected optokinetic reflex waveforms. The imaging device 504 captures images of the subject’s eye movement, which may be used to determine eye movement data including pupil motion and changes in the palpebral fissure height (PFH), as the subject reacts to the optokinetic stimuli. The captured video (e.g., input video signal) is transmitted to a processing device, where the data undergoes pre-processing. The processing device extracts eye movement waveforms, such as PFH versus time, and transforms the raw video data into a structured format suitable for further computational analysis.

[0248] The processed eye movement data is forwarded to a trained classification model for probabilistic analysis. The trained classification model, implemented within the processing pipeline / processing architecture, evaluates the extracted waveforms and eye movement features to determine patterns indicative of specific conditions. The trained classification model can apply machine learning algorithms trained on labeled datasets. The trained classification model can assign probabilistic scores to the observed patterns, identifying the likelihood of a given condition being present. For example, the trained classification model may analyze irregularities in saccadic resets or disruptions in PFH patterns to probabilistically determine whether a condition such as optokinetic dysreflexia or neurological impairment exists.

[0249] Reference is now made to FIG. 7, which illustrates another example of optokinetic stimulation and eye movement data capture 700. The stimulation may be processed by the user apparatus 500 of FIG. 5.

[0250] The display device 502 is configured to present vertically oriented, high-contrast optokinetic stimuli designed to elicit vertical nystagmus. The stimuli induce reflexive movements where the eyes alternately look upward and downward while the eyelids open and close. The motion provides for the measurement of palpebral fissure height (PFH) as a function of time, which provides for evaluating the optokinetic reflex (OKR). The vertical stimuli are calibrated to precisely stimulate the cortical and subcortical pathways involved in oculomotor control.

[0251] The imaging device 504 is mounted to record video of the subject’s eyes during the test. The imaging device can capture detailed movements of both the upper and lower eyelids and tracks pupil motion in response to the stimuli. The video data is transmitted to a processing device which extracts waveform data representing the PFH versus time for both eyes. The waveform captures the reflexive dynamics of the OKR and serves for subsequent probabilistic analysis.

[0252] The processing device can generate feature-rich outputs from the captured waveforms to include in the eye movement data. For example, the processing device may generate correlated waveforms from both eyes. By capturing vertical nystagmus with corresponding PFH measurements, the processing device may provide a quantifiable medium for further data processing.

[0253] Reference now first made to FIG. 8, which illustrates an example of optokinetic stimulation and eye movement data capture 800. The stimulation may be processed by the user apparatus 500 of FIG. 5.

[0254] The imaging device 504 is configured for measuring vertical optokinetic nystagmus to upward-moving stimuli. The imaging device 504 records video of both the upper and lower eyelids and pupil movements in response to upward-directed visual stimuli displayed on the display device 502. The upward stimulation elicits reflexive eye movements, including both upward tracking and periodic resetting motions, providing data for assessing vertical oculomotor function. Additionally, the imaging device 504 captures detailed palpebral fissure height (PFH) variations over time as the eyes reflexively respond to the stimuli.

[0255] The captured data includes multiple parameters, such as pupil displacement, eyelid dynamics, gaze direction, and velocity of eye movement. The imaging device 504transmits this raw video and extracted positional data to the processing device. The processing device is configured to process the data into waveform representations, including PFH versus time waveforms for the eye(s).

[0256] The processing device processes the refined waveform data and generates feature sets that are forwarded as eye movement data to the trained model selection module for further processing. The machine learning model performs a probabilistic analysis of the data to classify the observed eye movements against a range of predefined categories or conditions.

[0257] Reference is now made to FIG. 9, which illustrates an example optokinetic stimulation and eye movement data capture 900. The stimulation may be processed by the user apparatus 500 of FIG. 5.

[0258] The apparatus 500 is configured to measure horizontal optokinetic nystagmus (OKN) in response to a rightward-moving stimulus. The display device 502 presents dynamic horizontal stimuli such as alternating high-contrast stripes moving rightward across the subject’s field of view. The stimuli are designed to elicit a reflexive response characteristic of horizontal OKN, combining smooth pursuit eye movements tracking the stripes and rapid saccadic resets in the opposite direction. This setup provides that both “look” and “stare” phases of OKN are stimulated and measured.

[0259] The imaging device 504 captures real-time video of both eyes simultaneously as they move in unison to track the stimulus. The device 504 captures parameters such as pupil displacement, gaze direction, and saccadic velocity. The captured video data is then transmitted to the processing device for preprocessing and analysis. The processing device extracts one or more metrics, such as horizontal eye movement waveforms, and combines them into unified datasets representing both the “look” and “stare” components of the OKN reflex. The apparatus 500 is setup and the visual stimuli are displayed so that the left and right eyes of the subject are synchronized and moving in unison.

[0260] The processing device further refines the extracted waveforms to generate feature sets, such as amplitude and frequency of the horizontal OKN response. These features are then forwarded as eye movement data to a trained model selection module for further analysis and classification.

[0261] Reference is now made to FIG. 10, which illustrates an example optokinetic stimulation and eye movement data capture 1000. The stimulation may be processed by the user apparatus 500 of FIG. 5.

[0262] In an embodiment, the apparatus 500 is configured to measure optokinetic reflexes (OKR) in four directions i.e., rightward, leftward, upward, and downward, to produce an optokinetograph (OKG) that can detect optokinetic dysreflexia. The display device 502 presents dynamic visual stimuli, such as alternating high-contrast stripes, that move in the four cardinal directions. The directional stimuli are calibrated to elicit corresponding OKR responses including smooth pursuit movements in the direction of the stimulus followed by rapid saccadic resets in the opposite direction.

[0263] The imaging device 504 records real-time video data of both of the subject’s eyes simultaneously as they respond to the directional stimulus. The device 504 captures metrics, including the movement patterns, speed, and synchronization of the left and right eyes. For the stimulus direction, the imaging device 504 can track the characteristic waveform of the OKR response, which is then transmitted to the processing device. The processing device analyzes the raw video data to extract detailed waveforms, corresponding to the reflexive responses for the direction(s), as shown in the captured graphs for rightward, leftward, upward, and downward stimuli.

[0264] The processing device aggregates, packages, and formats the extracted directional waveforms into a structured dataset. The dataset includes features such as amplitude, frequency, and symmetry for the directional reflex. The processed data is then transmitted as eye movement data to the trained model selection module for further analysis.

[0265] In at least one embodiment, the apparatus 500 performs validity gating at the point of acquisition. The processing device suppresses logging, streaming, and / or persistence of eye movement data unless predetermined validity criteria are satisfied, including a valid gaze indication and valid bilateral pupil measurements from the imaging device 504. For a valid frame or interval, the processing device packages the eye movement data with protocol metadata identifying a stimulus block and epoch and the stimulus parameters used by the stimulus generator 506. The package provides for a structured session record without storing raw image frames. The structured sessionrecord may be streamed over the network 104 for real-time visualization at an administrator device 108 and. Upon session completion, the session record is uploaded and stored in external data storage 110 as a flat record (e.g., a JSON-formatted dataset) for offline or batch processing by the multi-model architecture 106.

[0266] In at least one embodiment, the multi-model architecture 106 implements a neural-network tagging layer that operates as a processing stage for generating an intermediate representation of the eye movement data prior to inference. The tagging layer may be implemented within the input layer 304 and / or as a pre-model stage executed by the processing device associated with the imaging device 504. The tagging layer is configured to receive raw eye movement data and / or pre-processed eye movement data derived from image data captured during presentation of visual stimuli. The tagging layer can implement a tagging routine trained on labeled datasets to segregate the received measurements into predefined tagged classifications. The tagging layer may assign discrete numerical or symbolic identifiers to a classified parameter and associates each tagged classification with the corresponding measured value(s) and timestamps. After tagging, the tagging layer performs a normalization process that scales continuous measurements into standardized numerical ranges configured for the selected classification model’s input requirements. The ranges may include any combination of scaling velocity, amplitude, and frequency measurements and encoding event-based features such as interruptions as discrete instances per interval. The output of the tagging layer is generated as classified data comprising normalized numerical values grouped under their corresponding tags and arranged as a structured dataset. The classified data provides an intermediate representation of the subject’s response that is transmitted to subsequent components such as feature extraction routines and the inference layer 308.

[0267] In at least one embodiment, condition-specific feature extraction is performed by the tagged classifications in the classified data rather than using a single fixed feature set across all trained classification models. The input layer 304 and / or associated feature extraction logic implements relevancy parameters comprising condition-specific tags to isolate subsets of the classified data for generation of feature data 306. The relevancy parameters can be determined during training of a trained classification model byprocessing labeled datasets that correlate tagged classifications with known outcomes for the condition. Further, for a model, a preset assessment tag set can be stored that identifies which tags are informative for that model. When the model identity is known or a target condition is specified, the input layer 304 applies the preset assessment tag set as a filter to select and extract the corresponding tagged classifications while excluding other tags to reduce processing volume. For example, for a dementia-related trained classification model, the preset assessment tag set may prioritize tags for pursuit gain, interruptions, look-to-stare ratio stability, and saccadic dynamics. The input layer 304 extracts the tagged values and their stimulus associations to generate feature data 306 for inference. For a stroke-related trained classification model, the preset assessment tag set may prioritize asymmetry-related tags, timing tags, and nystagmus irregularity tags. The input layer 304 can isolate those parameters while excluding unrelated smooth pursuit metrics. For vestibular or auditory pathway assessment, the preset assessment tag set can prioritize OKAN-related persistence features, off-axis nystagmus directionality and frequency, and stimulus-off phase duration metrics. In at least one embodiment, the preset assessment tag sets are further refined by adaptive learning mechanisms in which the trained classification model generates inferential data indicating probabilistic relevance of additional tags based on node weights, activations, and correlation matrices. The inferential data is stored as model configuration data that is applied as updated relevancy parameters during subsequent executions, such that feature extraction prioritizes newly identified tags (e.g., eyelid dynamics or PFA response time) without manual redefinition.

[0268] In at least one embodiment, the eye movement data processed by the multimodel architecture 106 includes, in addition to measured eye movement parameters, stimulus data. The stimulus data can be treated as a first-class input to the model selection module 204 and to the selected classification model 210. The stimulus data is generated by the stimulus generator 506 and includes explicit stimulus parameters captured as stimulus metadata, such as pattern type (e.g., alternating stripes, moving dot arrays, rotating fields), stimulus speed, stimulus direction, stimulus contrast and / or color configuration, and the testing environment settings including ambient light level, reduced light, or total darkness (e.g., for OKAN-related conditions). The stimulus metadata isstored and transmitted with a time-based association to the captured eye movement data as part of the input data 202. Therefore, a segment of eye movement data includes a corresponding stimulus identifier and the stimulus parameter values that elicited the segment. The model selection module 204 performs a probabilistic analysis that weights the stimulus metadata as selection parameters alongside the tagged classifications extracted from the eye movement data. For example, the model selection module 204 can apply a selection rule set in which vertical stripe directionality combined with a specified speed range and low ambient light condition increases a likelihood score for activating a trained classification model configured for PFA / PFH and Look-to-Stare Ratio processing. In another example, a high-speed horizontal stripe configuration with high contrast increases a likelihood score for a trained classification model configured for optokinetic dysreflexia (OKD) or horizontal OKN analysis. The selection signal 205 thereby implements stimulus-aware inference routing, in which routing is performed based on an evaluated combination of (i) stimulus metadata and (ii) tagged and normalized eye movement parameters. The selected classification model 210 can further process the stimulus metadata as contextual input during inference for scoring. In at least one embodiment, the inference layer 308 receives feature data 306 that includes the tagged classifications and the stimulus metadata. The layer 308 implements weights to one or both of the movement features and the stimulus parameters. Therefore, a pursuit gain value, interruption frequency, or PFA symmetry score is evaluated relative to the stimulus speed, direction, and ambient light state associated with the segment from which the feature was extracted.

[0269] In at least one embodiment, the plurality of visual stimuli generated may be configured not only to elicit optokinetic nystagmus (OKN) based on pupil motion, but also to elicit measurable eyelid dynamics such that palpebral fissure area (PFA) and / or palpebral fissure height (PFH) become relevant measured signal for classification. The visual stimuli may comprise vertically oriented, high-contrast patterns configured with a motion range and speed profile selected to induce repeated eyelid opening / closing dynamics synchronized with the reflexive eye movement response. The stimuli can lead to generation of PFH versus time waveforms and / or PFA waveforms extracted from the image data captured by the imaging device 504. In such embodiments, the processingdevice can perform frame-by-frame eyelid edge detection by applying segmentation computations to identify upper and lower eyelid boundaries, determine PFA and / or PFH for a time frame, and generate waveform features. The features can include one or more of (e.g., any combination of) eyelid separation amplitude, response time (e.g., onset latency from stimulus initiation to measurable PFH change), symmetry (e.g., comparative PFA / PFH between left and right eyes), and lid lag metrics (e.g., delayed upper eyelid movement relative to eye movement direction during downward gaze). The input layer 304 can assign tags to these eyelid-derived features and normalize the tagged classifications into classified data for subsequent feature extraction. Therefore, the stimulus objective is directed to produce eyelid / palpebral waveforms as a measurable proxy signal for condition detection and model inference, rather than relying only on classic OKN stimuli that primarily target eye position and velocity measurements from pupil tracking.

[0270] In at least one embodiment, the apparatus 500 may implement a dark / occluded stimulus protocol for optokinetic after-nystagmus (OKAN) capture in which a stimulus-off phase is a portion of the designed test. The display device 502 presents an optokinetic stimulus for a defined interval. The stimulus generator 506 terminates the stimulus at a defined timestamp while the visor 508 occludes external light to present darkness. During the stimulus-off phase, the imaging device 504 operates in an infrared (IR) capture mode to record continued eye / eyelid responses following cessation, and the processing device segments the captured video into “stimulus-on” and “stimulus-off” intervals. The processing device extracts tagged classifications during the stimulus-off interval. The tagged classifications may include one or more of sustained nystagmus frequency, decay rate, response duration, and PFA / PFH persistence metrics. The processing device associates tagged classifications with stimulus metadata identifying an ambient light state of “darkness” and a cessation marker. The input layer 304 can normalize the tagged classifications from the stimulus-off interval into classified data. The input layer 304 can generate feature data 306 for a trained classification model configured for vestibular and auditory pathway assessment.

[0271] In at least one embodiment, the user device 102 and the multi-model architecture 106 are configured to provide for optokinetic stimulus generation, validity-gated eye-tracking acquisition, and low-bandwidth feature formation that improves computational efficiency while maintaining reproducibility. For example, the user device 102 may be implemented as a head-mounted display device including integrated eye tracking executing a real-time application fully on-device during acquisition, without cloud dependency. The user device 102 renders a controlled optokinetic stimulus (e.g., a moving grating) parameterized by direction (e.g., left-to-right or right-to-left), angular speed (e.g., degrees / second), and phase (e.g., 0-1 continuous cycle), such that the stimulus rendering is deterministic and parameter-driven for repeatable test runs. During the OKR / OKN capture, the runtime may suppress logging, streaming, and / or downstream computation unless eye-tracking validity criteria are satisfied (e.g., gaze is valid and both pupils report non-zero valid diameters). The feature may reduce unusable samples at the source rather than relying on later filtering and reducing CPU, memory, and I / O utilization.

[0272] The multi-model architecture 106 introduces intelligence primarily at the interpretation and classification stages, while reducing reliance on raw pixel streams or high-frequency noisy sensor signals. The classification model(s) receive derived physiological features computed at or near the edge, such as gaze yaw angle, gaze angular velocity, saccade flags (e.g., thresholded velocity events), and bilateral pupil diameters, along with protocol metadata (e.g., block / epoch identifiers and stimulus parameters). The multi-model architecture 106 may additionally process block-level aggregates and reflex-dynamics features (e.g., slow-phase velocity distributions, saccade frequency / timing, directional asymmetry metrics, pupil symmetry and temporal variance) rather than raw frames. The feature-first design enables smaller and faster inference models (e.g., lightweight neural networks or temporal models rather than computeintensive video-based architectures). The feature reduces power consumption and provides predictable real-time performance on standalone hardware. In at least one embodiment, frames and / or feature records may be streamed over a local network for live visualization (e.g., LAN-based streaming to a local dashboard) and stored at session end as a structured record (e.g., a flat JSON session file uploaded to a session API).

[0273] In at least one embodiment, the user device 102 includes an imaging device 504 configured to capture image data of a subject’s eyes during presentation of the visual stimuli. In at least one embodiment, the user device 102 is configured to provide a validgaze determination during stimulus response acquisition. The user device 102 is configured to process eye-tracking validity signals generated by an eye-tracking interface associated with the imaging device 504. In at least one embodiment, the user device 102 is configured to receive, per frame or per short acquisition interval, a combined gaze validity field that may be adapted from an eye-tracking software development kit (SDK) associated with the imaging device 504. The user device 102 can determine a gaze-valid state based on the combined gaze validity field. For example, in at least one embodiment the user device 102 can determine bilateral pupil validity by reading a left pupil diameter field and a right pupil diameter field and requiring presence and non-zero values for both fields in the same frame. The user device 102 can analyze numeric integrity and bounds for gaze angle values and pupil diameter values by rejecting frames associated with non-finite values, such as not-a-number (NaN) values or infinite (Inf) values. The user device 102 can analyze numeric integrity and bounds for gaze angle values and pupil diameter values by rejecting frames associated with values outside stored sanity ranges maintained for the test configuration. The user device 102 can process one or more tracking-quality indicators determined using eye-tracking (e.g., adapted from the SDK), such as one or more of a confidence value, a tracking state value, or a quality state value. The user device 102 can determine whether the indicator satisfies an acceptability criterion stored for the acquisition session. The user device 102 can determine an implicit blink or occlusion state without eyelid image processing by treating missing, invalid, or zeroed pupil diameter fields as an occlusion indicator for the frame. The user device 102 can cause the frame to fail the bilateral pupil validity determination based on the aforementioned determinations.

[0274] In at least one embodiment, the user device 102 may analyze head pose parameters generated by the eye-tracking interface and apply defensive head pose sanity bounds to reject frames associated with extreme pose values, which may result the measurements leading to inadequate / inaccurate testing under certain circumstances. The user device 102 can maintain the head pose sanity bounds as a reject-only criterion rather than a primary gating signal. The user device 102 can generate a validity-gated eye movement data stream for downstream tagging, normalization, feature extraction,and inference in the multi-model architecture 106. The functional purpose includes reducing propagation of invalid gaze samples into classified data and feature data 306.

[0275] In at least one embodiment, the user device 102 is configured to perform validity thresholding and time windowing at an acquisition rate associated with the imaging device 504 and an eye-tracking interface coupled to the imaging device 504. The user device 102 can process, for a frame, or several frames for a short acquisition interval, a set of validity fields and sanity checks and generate a per-frame validity decision in real time. The user device 102 can access stored gating parameters in memory, including one or more of confidence thresholds, numeric sanity ranges, or reject-only head pose bounds. The user device 102 can apply the gating parameters to the received per-frame values to determine a valid frame state or an invalid frame state. The user device 102 can associate a valid frame state with a valid gaze sample record comprising gaze angle values, bilateral pupil diameter values, and / or validity metadata. The user device 102 can associate the valid gaze sample record with a timestamp aligned to stimulus metadata generated by the stimulus generator 506.

[0276] In at least one embodiment, the user device 102 may be configured to determine that a frame is valid when the required validity conditions are concurrently satisfied within the same frame interval, including a combined gaze-valid flag state, bilateral pupil validity, numeric integrity checks, and any enabled tracking-quality criteria. The user device 102 can determine that a frame is invalid when one or more required conditions fail, including transient invalid states associated with blinks or occlusions inferred from missing or invalid pupil diameter values. The user device 102 can generate a data stream comprising valid gaze sample records and permit gaps in the data stream corresponding to transient invalid states. The process can be performed without generating interpolated samples and without generating synthetic back-filled values. The user device 102 can accept a valid frame without enforcing a minimum dwell time. The user device 102 can forward the validity-gated stream as eye movement data to the multi-model architecture 106 for downstream block-level or epoch-level computational processing. The processing can include one or more of aggregation routines, smoothing stages, or feature extraction operations performed by the input layer 304.

[0277] In at least one embodiment, the user device 102 is configured to control acquisition-side handling when a frame is determined to be invalid. The user device 102 can execute on the processing device associated with the imaging device 504 to suppress invalid frames at the source prior to generation of eye movement data transmitted over the network 104. The user device 102 can implement a write-path gate for one or more data sinks, such as one or more of a local session buffer, a streaming output interface, or a persistence interface coupled to external data storage 110. The user device 102 can, upon determining an invalid frame state, block transfer of the corresponding frame measurements to the data sinks such that the invalid frame measurements are not logged, are not streamed, and are not stored. The user device 102 can maintain a strict non-synthesis rule in which no synthetic samples are generated for an invalid frame and no back-filling values are computed for the invalid frame interval. The user device 102 can forward valid gaze sample records for downstream tagging and feature extraction by the input layer 304.

[0278] In at least one embodiment, the user device 102 includes a stimulus generator 506 configured to continue rendering the plurality of visual stimuli on the display device 502 without interruption during invalid frame intervals. The user device 102 can maintain an invalid-duration counter and determine an extended invalid period when consecutive invalid frames exceed a stored duration threshold or count threshold. The user device 102 can generate runtime diagnostics data indicating the extended invalid period. The runtime diagnostics data may include one or more of a timestamp range, a counter value, a current tracking-quality indicator state, or a head pose reject indicator. The runtime diagnostics data may be stored as administration data and / or transmitted to the administrator system 108 for display. The user device 102 can perform a recalibration operation or a restart operation to restrict automatic recalibration triggers during acquisition to maintain deterministic stimulus timing and auditability of the session record. The functional purpose of invalid-frame suppression is to prevent propagation of invalid measurements into the eye movement data stream and into the downstream trained classification model processing.

[0279] In at least one embodiment, the user device 102 is configured to generate a session record from validity-gated eye movement data during presentation of visualstimuli by the stimulus generator 506. The user device 102 can receive valid gaze sample records and associated stimulus metadata and to store the received records to external data storage 110 and / or local memory of the user device 102 as a structured session record. The user device 102 can exclude invalid frames by construction, based on the write-path gate, thereby limiting the session record to samples that satisfy concurrently evaluated validity conditions. The user device 102 can store the valid gaze sample records with temporal markers, including timestamps and block or epoch identifiers, such that later processing can determine sample provenance relative to stimulus configuration and acquisition timing. The user device 102 further can store validity metadata for the accepted samples, including one or more of combined gaze-valid flag states, trackingquality indicator values, or reject reason indicators.

[0280] In at least one embodiment, the user device 102 may be configured to maintain gaps in the recorded time series for intervals in which the user device 102 determines invalid frame states by suppressing invalid frames prior to session record storage. The gaps are configured to correspond to headset-level tracking loss events rather than to post-hoc removal of stored samples. The multi-model architecture 106 can receive the session record or a streamed subset of the session record as input data 202 and process the validity-gated dataset without requiring access to raw image frames captured by the imaging device 504. The input layer 304 can process the validity-gated samples to generate classified data and feature data 306. The process can include applying normalization routines and condition-specific tag filters at block or epoch scope. The inference layer 308 can process the feature data 306 to generate evaluation data and assessment data. The functional purpose of the validity-gated session record is to provide a reduced dataset that preserves physiologically meaningful measurements while maintaining explicit tracking-loss gaps as part of the session representation. The functional purpose further includes enabling downstream inference to operate on derived eye movement data and stimulus metadata without dependence on raw pixel streams.

[0281] In at least one embodiment, the user device 102 may be configured to generate a gaze heatmap during stimulus response acquisition for pupil-based measurements and for gaze validity checking. The user device 102 can receive, for a frame interval, image-derived eye feature data and / or eye-tracking interface outputs associated with theimaging device 504. The user device 102 can process the received inputs to generate a heatmap representation that encodes a gaze location in a coordinate space aligned to the display device 502 and the presented visual stimulus. The user device 102 can determine a pupil’s location in a frame. The user device 102 can map the pupil location to a gaze point and generate a spatial intensity distribution centered on the gaze point. In at least one embodiment, the user device 102 can associate the heatmap representation with stimulus metadata, such as one or more of a stimulus direction, a stimulus pattern type, and / or a stimulus block identifier. The user device 102 can analyze the heatmap representation to generate gaze deviation metrics, such as one or more of a distance-to-stimulus-region metric, an off-axis gaze angle metric, or an extreme gaze indicator associated with upward, downward, or lateral gaze beyond stored bounds. The user device 102 can include the gaze deviation metrics in the eye movement data and / or as auxiliary features processed by the input layer 304.

[0282] In at least one embodiment, the user device 102 may be configured to process the heatmap representation to generate pupil size measurements and pupil motion measurements. The user device 102 can determine a pupil diameter value by estimating a pupil region extent based on a heatmap of the obtained eye data such as that associated with a heatmap peak, for example. The user device 102 can convert the pupil region extent into a pupil size metric in the coordinate space. The user device 102 can determine pupil movement by tracking a sequence of heatmap peak locations over time. The user device 102 can generate one or more motion features, such as one or more of displacement values, velocity values, or directionality values. In at least one embodiment, the user device 102 can determine an eye-closed state by detecting an absence of a pupil region in the inputs for the frame interval. The user device 102 can generate a heatmap output corresponding to a null or zero measurement state for the frame interval. The user device 102 can determine a pupil validity failure for the frame based on the null or zero measurement state. The user device 102 can process the heatmap-derived metrics as an additional sanity check path. The user device 102 can reject frames associated with extreme gaze indicators that correspond to avoidance of a presented stimulus region. The functional purpose of heatmap-based pupil measurements and sanity checks is to generate gaze-location and pupil measurement signals suitable for acquisition-timevalidity gating and protocol adherence. The functional purpose further includes detecting eye closure and extreme off-stimulus gaze conditions during acquisition-time quality control.

[0283] In at least one embodiment, the user device 102 may be configured to determine whether an optokinetic reflex (OKR) test segment satisfies a duration criterion for acceptance. The user device 102 can process validity-gated samples and maintain a cumulative valid-duration value for a stimulus block identifier or epoch identifier generated by the stimulus generator 506. The user device 102 can generate a block acceptance signal when the cumulative valid-duration value meets a stored time limit for the block. The user device 102 can generate a block rejection signal when the stored time limit is not met by a termination event for the block. In at least one embodiment, the user device 102 is configured to maintain the stored time limit as a configurable parameter rather than a fixed value. The user device 102 can select the stored time limit from a range of candidate time limits determined by experimental testing on a population. The administrator system 108 and / or the multi-model architecture 106 can analyze collected session records to determine an updated time limit, including condition-specific time limits or stimulus-specific time limits. The user device 102 can store the updated time limit for subsequent sessions. The user device 102 can continue stimulus rendering and continue accumulating valid-duration when valid gaze sample records satisfy gating criteria, including during extended segments associated with delayed stimulus effects.

[0284] In at least one embodiment, the user device 102 is configured to analyse pupil diameter validity during acquisition. The user device 102 can determine a pupil diameter value from one or more pupil measurement fields, such as one or more of a left pupil diameter field or a right pupil diameter field. The user device 102 can compare the pupil diameter value to a lower bound and an upper bound stored for a test configuration. The user device 102 can set the upper bound relative to an iris size estimate derived from the same eye feature data. The upper bound can correspond to a value approaching an iris extent without requiring a fixed physical unit. In at least one embodiment, the user device 102 is configured to analyze head position validity during acquisition. The user device 102 can receive inertial sensor data, including data from one or more of a gyroscope or an accelerometer. The user device 102 can determine head pose parameters, such asone or more of pitch values, roll values, or yaw values. The user device 102 can compare the head pose parameters to stored position limits and to generate a head pose accept indicator or a head pose reject indicator for a frame interval. The user device 102 can treat the head pose reject indicator as a defensive gating input when determining validity for the frame interval. The user device 102 can update the stored position limits based on experimental population testing, including testing that correlates excessive tilt to degraded reflex measurements. The functional purpose of pupil bounds checks and head pose bounds checks is to provide configurable acquisition acceptance criteria that can be tuned through population testing. The functional purpose further includes limiting invalid pupil measurements and excessive head pose deviations that can degrade optokinetic stimulus-response measurements.

[0285] The multi-model architecture 106 is further configured to provide adaptive, Al-driven protocol optimization that reduces rendering time and acquisition burden while preserving or improving decision confidence. For example, the architecture 106 may detect early convergence of reflex metrics, shorten or skip redundant stimulus blocks, and / or adjust stimulus parameters (e.g., stimulus speed or duration) dynamically based on intermediate assessment outputs and confidence metrics for data validity. Such closed-loop coordination between stimulus presentation, feature extraction, and model inference attributes the observed efficiency improvements to the disclosed configuration. These improvements include reducing total rendered stimulus time, decreasing the number of samples collected and processed, and avoiding downstream computation on invalid data.

[0286] The visual testing described in the various embodiments herein may be used in a variety of different situations. For example, a sports player that received an impact which may have caused a concussion may be tested using one of the embodiments described herein to determine if the player really did receive a concussion and optionally an assessment of the severity of the concussion. In another example, a law enforcement professional who suspects that a driver is intoxicated or have taken mind-altering drugs may use one of the embodiments described herein to test whether the driver is sober or under the influence of alcohol or drugs. Additional applications include concussion screening, fatigue monitoring, cognitive assessment, and intoxication detection. The usecases can be applied by various professionals (e.g., sports officials, law enforcement, or employers in high-risk industries) in assessing neurological reflex impairments.

[0287] It should be noted that various references to a user device performing certain functions may be implemented using one or more processors of the user device to execute software instructions for performing the functions.

[0288] Numerous specific details are set forth herein in order to provide a thorough understanding of the exemplary embodiments described herein. However, it will be understood by those of ordinary skill in the art that these embodiments may be practiced without these specific details. In other instances, well-known methods, procedures and components have not been described in detail so as not to obscure the description of the embodiments. Furthermore, this description is not to be considered as limiting the scope of these embodiments in any way, but rather as merely describing the implementation of these various embodiments.

Claims

CLAIMS1. A device for feature extraction and assessment of a condition of a subject by performing visual testing on the subject, the device comprising:a processor configured to:generate a plurality of visual stimuli configured to elicit eye movement in a subject, the plurality of visual stimuli comprising one or more patterns;display the plurality of visual stimuli to the subject;receive image data capturing any eye movements and any eye lid movements of the subject due to presentation of the plurality of visual stimuli where the image data comprises optokinetic reflex (OKR) responses;process the image data to obtain eye movement data including at least one measurement of eye movement and / or at least one measurement of eye lid movement;select a trained classification model by performing a probabilistic analysis on a set of selection parameters applied to the eye movement data, the set of selection parameters comprising characteristics of a plurality of eye movement data;segregate the plurality of eye movement data to generate tagged classifications by assigning predefined tags to measurable parameters of the eye movement data and converting the tagged classifications into normalized numerical values to generate classified data;generate feature data by extracting the tagged classifications from the classified data based on relevancy parameters associated with the trained classification model, the relevancy parameters comprising condition-specific tags; anddetermine or assess the condition based on a probability score.

2. The device of claim 1 wherein the processor is configured to generate the probability score using by processing the tagged classifications in the feature data with assigned weights to identify patterns indicative of the condition and optionally comparing the probability score to a threshold.

3. The device of claim 1 or claim 2, wherein the processor is configured to generate administration data for the condition based on a pre-defined database configured with treatment protocols specific to the condition, the administration data comprising treatment recommendations.

4. The device of claim 3, wherein the processor is configured to provide the treatment recommendations based on the determined or assessed condition to a person who performed the visual testing.

5. The device of any one of claims 1 to 4, wherein the plurality of visual stimuli is generated on a display screen and the image data is captured by an image sensing device, and wherein the one or more patterns may have one or more speeds, one or more directions, and one or more colors.

6. The device of any one of claims 1 to 5, wherein the eye movement data comprises any combination of amplitude, frequency, velocity, pursuit gain, interruptions, symmetry, and directionality of eye movements.

7. The device of any one of claims 1 to 6, wherein the selection parameters provide an indication of a potential condition, and the selection parameters include any combination of saccadic movements, smooth pursuit, reflexive gaze stabilization, pursuit gain, and interruption frequency, and wherein the selection parameters are determined by user input.

8. The device of any one of claims 1 to 7, wherein the predefined tags comprise any combination of nystagmus, saccades, smooth pursuit, look-to-stare ratios, and palpebral fissure area (PFA) measurements.

9. The device of any one of claims 1 to 8, wherein the relevancy parameters are identified by performing a probabilistic analysis on training data comprising labeled datasets, and wherein the probabilistic analysis involves correlating the relevancy parameters with the condition.

10. The device of any one of claims 1 to 9, wherein the processor is configured to generate inferential data identifying a plurality of additional condition-specific tags based on the analysis of node weights and correlation patterns by the neural network, the inferential data providing probabilistic relevance of untagged parameters to the condition.

11. The device of any one of claims 1 to 10, wherein the condition comprises a neurological condition or a physiological condition wherein the neurological condition includes dementia, Parkinson’s disease, vestibular disorders, cranial nerve palsies and stroke, and wherein the physiological condition includes intoxication and fatigue.

12. The device of any one of claims 1 to 11, wherein the processor is configured to perform validity gating for suppressing logging, streaming, and / or persistence of eye movement data unless predetermined validity criteria are satisfied including a valid gaze indication and valid bilateral pupil measurements.

13. A method for feature extraction and assessment of a condition of a subject by performing visual testing on the subject, the method comprising:generating a plurality of visual stimuli configured to elicit eye movement in a subject, the plurality of visual stimuli comprising one or more patterns;displaying the plurality of visual stimuli to a subject via a display;receiving image data capturing any eye movements and any eye lid movements of the subject due to presentation of the visual stimuli where the image data comprises optokinetic reflex (OKR) responses;processing the image data to obtain eye movement data including at least one measurement of eye movement and / or at least one measurement of eye lid movement;selecting a trained classification model by performing a probabilistic analysis on a set of selection parameters applied to the eye movement data, the set of selection parameters comprising characteristics of a plurality of eye movement data;segregating the plurality of eye movement data to generate tagged classifications by assigning predefined tags to measurable parameters of the eye movement data and converting the tagged classifications into normalized numerical values to generate classified data;generating feature data by extracting the tagged classifications from the classified data based on relevancy parameters associated with the trained classification model, the relevancy parameters comprising condition-specific tags; anddetermining or assessing the condition based on a probability score.

14. The method of claim 13, wherein the method comprises generating the probability score by processing the tagged classifications in the feature data with assigned weights to identify patterns indicative of the condition and optionally comparing the probability score to a threshold.

15. The method of claim 13 or claim 14, wherein the method further comprises generating administration data for the condition based on a pre-defined database configured with treatment protocols specific to the condition, the administration data comprising treatment recommendations.

16. The method of claim 15, wherein the method including providing the treatment recommendations based on the determined or assessed condition to a person who performed the visual testing.

17. The method of any one of claims 13 to 16, wherein the plurality of visual stimuli is generated on a display screen of an electronic device including a monitor, a visor or a VR headset and the plurality of image data is captured by an image sensing device, and wherein the one or more patterns may have one or more speeds, one or more directions, and one or more colors.

18. The method of any one of claims 13 to 17, wherein the eye movement data comprises any combination of amplitude, frequency, velocity, pursuit gain, interruptions, symmetry, and directionality of eye movements.

19. The method of any one of claims 13 to 18, wherein the selection parameters provide an indication of a potential condition, and the selection parameters include any combination of saccadic movements, smooth pursuit, reflexive gaze stabilization, pursuit gain, interruption frequency.

20. The method of any one of claims 13 to 19, wherein the predefined tags comprise any combination of nystagmus, saccades, smooth pursuit, look-to-stare ratios, and palpebral fissure area (PFA) measurements.

21. The method of any one of claims 13 to 20, wherein the relevancy parameters are identified by performing a probabilistic analysis on training data comprising labeled datasets, and wherein the probabilistic analysis involves correlating the relevancy parameters with the condition.

22. The method of any one of claims 13 to 21, wherein the method further comprises generating inferential data identifying a plurality of additional condition-specific tags based on the analysis of node weights and correlation patterns by the neural network, the inferential data providing probabilistic relevance of untagged parameters to the condition.

23. The method of any one of claims to 13 to 22, wherein the condition comprises a neurological condition or a physiological condition wherein the neurological condition includes dementia, vestibular disorders, cranial nerve palsies and stroke, and wherein the physiological condition includes intoxication and fatigue.

24. The method of any one of claims 13 to 23, wherein the method further comprises performing validity gating for suppressing logging, streaming, and / or persistence of eye movement data unless predetermined validity criteria are satisfied including a valid gaze indication and valid bilateral pupil measurements24. A device for detection and / or assessment of a condition of a subject via visual testing, the device comprising:an image sensor;a display;memory that stores a trained Artificial Intelligence (Al) model; anda processor that is communicatively coupled to the image sensor, the display and the memory, wherein the processor is configured to:generate a plurality of visual stimuli configured to elicit optokinetic reflex (OKR) responses in the subject;display the plurality of visual stimuli to the subject;receive image data capturing any eye movements and any eye lid movements of the subject due to presentation of the plurality of visual stimuli; process the image data to obtain eye movement data including at least one measurement of eye movement and / or at least one measurement of eye lid movement;process the image data using the trained Al model to generate a detected condition assessment and / or a condition assessment of the detected condition; andprovide an indication of the detected condition assessment and / or the condition assessment of the detected condition.

25. The device of claim 24, wherein the measurements include a look-to-stare ratio, and / or palpebral fissure area (PFA) measurements.

26. The device of claim 24 or claim 25, wherein the Al model is one of a plurality of trained classification models that are associated with and trained for detecting and / or assessing a unique condition and one of the trained classification models is used as the Al model.

27. The device of claim 26, wherein the processor is configured to use trained selection model to perform a probabilistic assessment of the eye movement data to select one of the trained classification models to process the image data.

28. The device of any one of claims 24 to 27, wherein the processor is configured to provide treatment recommendations based on the detected condition assessment and / or the longitudinal tracking of the detected condition.

29. The device of any one of claims 24 to 28, wherein the condition comprises a neurological condition or a physiological condition wherein the neurological condition includes dementia, vestibular disorders, cranial nerve palsies and stroke, and wherein the physiological condition includes intoxication and fatigue.

31. The device of any one of claims 24 to 29, wherein the processor is configured to perform validity gating for suppressing logging, streaming, and / or persistence of eyemovement data unless predetermined validity criteria are satisfied including a valid gaze indication and valid bilateral pupil measurements30. A method for detection and / or assessment of a condition of a subject via visual testing, the method comprising:generating a plurality of visual stimuli configured to elicit optokinetic reflex (OKR) responses in the subject;displaying the plurality of visual stimuli to the subject;receiving image data capturing any eye movements and any eye lid movements of the subject due to presentation of the plurality of visual stimuli;processing the image data to obtain eye movement data including at least one measurement of eye movement and / or at least one measurement of eye lid movement;processing the image data using the trained Al model to generate a detected condition assessment and / or a condition assessment of the detected condition; and providing an indication of the detected condition assessment and / or the condition assessment of the detected condition.

31. A non-transitory computer-readable storage medium storing program instructions that, when executed by a processor, cause the processor to perform a method defined according to any one of claims 13 to 24 or 30.

Images