Pupillometry system for computer-implemented assessment of presymptomatic alzheimer's risk in users

Abstract

A pupillometry system (100) for computer-implemented assessment of brain aging and presymptomatic Alzheimer's risk in cognitively unimpaired individuals through computer-human interaction with a user (5), wherein the pupillary moment-by-moment dynamics of the user on a sensory stimulation decision task with continuous or—at least two—discrete selection options and a subsequent reward / feedback choice is analyzed.

Description

[0001] ETH 2023-172

[0002] 1

[0003] TITLE

[0004] PUPILLOMETRY SYSTEM FOR COMPUTER-IMPLEMENTED ASSESSMENT OF PRESYMPTOM ATIC ALZHEIMER'S RISK IN USERS

[0005] TECHNICAL FIELD

[0006] The present invention relates to a pupillometry system for computer-implemented assessment of brain aging and presymptomatic Alzheimer's risk in users through computerhuman interaction with a user.

[0007] PRIOR ART

[0008] As longevity expands globally due to medical advancements, a surge in age-related neurological conditions, particularly neurodegenerative diseases involving dementia, has become a prominent health concern. Alzheimer's Disease (AD), the predominant cause of dementia, affects approximately one in nine individuals aged 65 and above in the US, contributing to a global count exceeding 55 million in 2023. The late stages of AD involve significant healthcare expenditures primarily due to caregiving responsibilities. A startling fact is that over 70% of AD instances go unrecognized globally (most pronouncedly in lower income countries), and typically, diagnosis occurs post manifestation of cognitive symptoms. Present diagnostic methodologies once symptoms are evident involve, neuroimaging, and cerebrospinal fluid analysis, each incurring substantial costs and are not accessible to most people. By the time symptoms are evident, approximately two decades of pre-symptomatic disease progression, causing irreversible neural damage, have usually passed. At this stage the brain cannot be rescued.

[0009] Clinical diagnosis of AD nowadays occurs only after cognitive symptoms are detectable and require physical, neurocognitive and neurological examination, structural brain imaging (MRI or CT), and cerebrospinal fluid tests or PET scans (amyloid-p and tau protein markers), which require traveling to specialized clinical centers4. Critically, these tests are time-intensive and costly, making timely screening inaccessible for many individuals, given ETH 2023-172

[0010] 2 that underdiagnosis is most prevalent in low- and middle-income countries.

[0011] Additionally, there is currently no solution / system published or available to reliably monitor and predict AD risk in cognitively unimpaired individuals, alongside the continual assessment of risk evolution in decentralized settings.

[0012] WO2015121317A1 relates to a method to predict whether a male subject has an increased risk for Alzheimer's disease based on determining a fraction of cells from a biological sample of said male subject that have lost chromosome Y, comparing said fraction of cells to a predefined threshold and making said decision based on the result of the comparison.

[0013] WO2021161318A1 discloses a system, device and method for determining and / or assessing brain related conditions based on pupil light response (PLR) starting from determining a baseline pupil size of an eye of the subject, applying blue and / or red stimuli to one or more regions of the visual field of the eye, obtaining a value for one or more parameters related to induced changes in the pupil size in response to the light stimuli and classifying the normalized results in monitoring progression of, determining and / or assessing the brain related condition.

[0014] WO2021165193A1 discloses a method and assembly for determining a representation which indicates the activity of a locus coeruleus system of a user.

[0015] WO2018017767A1 discloses example systems, methods, and apparatus, including cognitive platforms, for applying signal detection metrics in computer-implemented adaptive response-deadline procedures to data collected based at least in part on user interaction(s) with computerized tasks and / or interferences. The apparatus can include a response classifier for generating a quantifier of the cognitive abilities of an individual.

[0016] US2014282646A1 discloses a method of monitoring the eyes of a viewer in real-time when this viewer is viewing three-dimensional contents on a consumer electronic device. Then biometric data from the monitoring can be used for delivery of highly accurate and personalized viewing experiences for the viewer.

[0017] Hossain Gahangir et al. have explained in the article "When does an easy task become hard? A systematic review of human task-evoked pupillary dynamics versus cognitive efforts", published in Neural Computing and Applications, Springer London, London, vol. ETH 2023-172

[0018] 3

[0019] 30, no. 1 , 17.12.2016, pages 29 to 43, XP036521894, ISSN: 0941-0643, DOI: 10.1007 / S00521-016-2750-5.

[0020] Granholm Eric L. et al. have provided an overview on "Pupillary Responses as a Biomarker of Early Risk for Alzheimer's Disease", in Journal of Alzheimer's Disease, vol. 56, no. 4, 20- 02-2017, pages 1419-1428, XP093247481 , ISSN: 1387-2877, DOI: 10.3233 / JAD-161078. In a similar context, Zeeman Michael et. Al. have provided a review on "Task-Evoked Pupillary Response as a Potential Biomarker of Dementia and Mild Cognitive Impairment: A Scoping Review", in American Journal of Alzheimer's Disease and other Dementias, vol. 38, 10-03-2023, XP093247492, ISSN: 1533-3175, DOI: 10.1177 / 15333175231160010.

[0021] SUMMARY OF THE INVENTION

[0022] Given the absence of effective curative treatments for Alzheimer's Disease (AD), the present invention has as one object to provide non-invasive, portable, and accurate biomarkers for early, pre-symptomatic prediction of AD, decades before cognitive symptom onset.

[0023] The invention is related to proposing a system using pupillometry and preferably causal inference for continual assessment of brain aging and presymptomatic Alzheimer's risk in individuals through computer-human interaction. It is an advantage of the present system according to the invention that it can be used to monitor cognitively unimpaired individuals.

[0024] The invention provides a pupillometry system for computer implemented assessment of brain aging and presymptomatic Alzheimer's risk in users through computer-human interaction with a user with the features of claim 1 .

[0025] The invention is based on the insight that pupil dynamics, especially relating to pupil size, based on decision policy learning allows a prediction before disease onset and when users are still cognitively unimpaired.

[0026] The invention addresses inter alia some problems of the prior art methods as:

[0027] - Predictive methods for AD and biological aging are currently dependent on laboratory evaluations conducted in specialized facilities.

[0028] - These procedures are cost-prohibitive, and regular monitoring for AD risk or accelerated ETH 2023-172

[0029] 4 aging is not economically viable.

[0030] - Consequently, there is a lack of reliable methods for ongoing monitoring of AD progression and biological aging.

[0031] - Further complicating the issue is the absence of integrated systems that combine methods, apparatus, and data for accurate calibration of such assessments.

[0032] - Additionally, there is a significant gap in our knowledge regarding the identification of users as individuals who will develop AD or exhibit signs of accelerated biological aging while still cognitively unimpaired.

[0033] The invention provides a pupillometry-based and causal inference system to generate continual assessments of brain aging and presymptomatic risk of Alzheimer's disease in cognitively unimpaired users based on computer-human interactions of decision-making behavior using pupil dilation dynamics.

[0034] The use of pupillometry within the framework of the invention is based on the insight that:

[0035] (i) locus coeruleus (LC) is one of the areas earliest affected in AD (structural level, e.g. MRI based),

[0036] (ii) pupil size is an accurate estimation of functional LC integrity, and

[0037] (iii) LC function is challenged and essential in situations that involve uncertainty (e.g. decision-making)

[0038] The invention is based on the use of a sensor to track the pupil of at least one eye while humans engage in decision-making tasks involving decision policy learning (or forming abstract cognitive maps), so it is possible to exploit the moment-by-moment fluctuations of pupil dilation dynamics (at the millisecond resolution) to predict brain aging (as opposed to chronological age) and presymptomatic risk of Alzheimer's disease in cognitively unimpaired individuals. Of course, utilizing a sensor to track the pupil of the two eyes simultaneously is preferred.

[0039] The calibration of the method and apparatus of such markers can be based on the unique data acquisition of a large cohort of cognitively unimpaired individuals of different groups which included the presenilin-1 (PSEN1) E280A mutation for autosomal dominant AD and non-carrier family members, alongside age-matched non-family members and healthy nonfamily members from a different continent and socio-economic levels. Also, young people from different continents would not be specific here. ETH 2023-172

[0040] 5

[0041] The invention provides several advantages over the prior art, e.g.:

[0042] - Earlier risk identification of AD since prior art systems and methods refer to cognitive assessments of AD based on sensitive cognitive tests, but it is at a stage of cognitive decline (e.g. MCI, extensive brain damage has already occurred) which is also due to lack of evidence on cognitively unimpaired individuals with high degrees of accuracy.

[0043] - Continuous, accessible and accurate assessment of brain age and AD risk is possible: In contrast to blood, MRI or imaging-based technologies, hardware as disclosed in the application of the system and method according to the present invention can be deployed also in rural regions and used to provide long-term monitoring of biomarker development in real-time allowing for the evaluation of temporal evolvement (e.g. speed of decline or improvement e.g. in combination with lifestyle and intervention)

[0044] - The system and method according to the invention uses an eye-tracking sensor and a computer-implemented method comprising easy-to-understand economic games with decision policy learning that extracts brain function information of brain areas involved in silent neuropathological damage via dynamic fluctuations of pupil dilation. This measurement does not necessarily require a highly controlled measurement environment (e.g., a person sitting in any room with a laptop / tablet with an integrated / attached eyetracker synchronized with the task) to predict with a high degree of accuracy who is at risk of developing Alzheimer’s disease in the future (i.e. , at presymptomatic stages when no other standardized method can accurately predict Alzheimer’s disease).

[0045] - Thus, a key ingredient of the invention lies in a causal inference model that extracts from pupil fluctuations improved decision policy learning strategies related to individuals’ brain function. This information is fundamentally revealing of AD risk condition, which is currently calibrated with data from presymptomatic (i.e., cognitively unimpaired) individuals that are certain to develop Alzheimer’s disease, alongside control groups to predict the gradual process of brain aging and the rate at which this occurs (i.e., more or less accelerated in certain individuals); this, alongside providing accurate risk estimation of developing AD in the future. The causal inference model can also be calibrated by other approaches, e.g. longitudinal study of a cohort of individuals performing the decision task at different points in time.

[0046] Within the stimuli selection, participants or users are presented with at least two stimuli that can be visual, auditory, tactile, olfactory, or gustatory. Relating to the trial count, each assessment consists of multiple trials, preferably approximately 50-100 trials. The reward structure comprises that the task is designed with a hidden reward pattern to encourage the exploration and strategic exploitation of the stimuli for maximum reward. The response ETH 2023-172

[0047] 6 mechanism can comprise different input units for the user.: Participants select options using an interface, which may include buttons or provide feedback visually, tactically, or acoustically. The pupil racking comprises monitoring pupil size during the task using a stationary or portable system, capturing data from either one or both eyes.

[0048] An acoustic presentation of selections can comprise different frequencies, left and right ear distribution of the sound, type of sound, etc.

[0049] A further alternative is to deliver the value of the hidden reward in an auditory manner via frequency tones or volume. For instance, if there are N=2 bandits, and the human participant uses her left hand to indicate that she has chosen bandit A, then the hidden reward of the chosen bandit can be revealed via a low-volume tone indicating a low reward value (or a high-volume tone for a high reward value) delivered over the left ear. Conversely, if the human participant uses her right hand to indicate that she has chosen bandit B, then the tones are instead delivered over the right ear.

[0050] A similar principle can be used in delivering information about the reward of the bandits using tactile feedback. Here, instead of using auditory tone volume, one could, for instance, the amount of mechanical pressure applied over the left or right index finger. Alternatively, one could also use mechanical frequencies via microactuators.

[0051] The system engages users, i.e. human participants, in learning cognitive map spaces while transitioning from exploration and exploitation states while making decisions. One such way is the utilization of a family of tasks known in the decision-making literature as “bandit tasks”. This is the family of tasks that is described as a specific embodiment, but any other task that induces the generation of cognitive map spaces (or also spatial maps, temporal maps, etc) lead to similar outcomes.

[0052] In the bandit tasks, the goal is to collect as much reward as possible by selecting the “bandit” (or selection option) that promises more reward out of N bandits according to the learned reward history. In the present task, we use N=2 bandits (but more are possible) in which participants attempt to learn the bandit rewards over T=100 trials (more trials are of course possible, but this amount of trials keeps participants engaged while being sufficient to achieve acceptable levels of accuracy). The true reward of each bandit is hidden, but only the one that is chosen is revealed in each trial (the reward of the remaining unchosen bandits is not revealed). In the present task, the value of the rewards is visually revealed ETH 2023-172

[0053] 7 using a computer / tablet / mobile-phone screen, however, as explained below, the value of the reward can be revealed using any other sensory modality (e.g., tones, tactile). A key aspect here is that the hidden rewards of each bandit drift following a random walk path. This random walk is governed by a stochastic equation. One such stochastic equation can be for instance one that makes the rewards of the bandits drift up and down around a mean reward. However, it is to be emphasized that the specific form of these stochastic transitions is not necessarily relevant, but is rather selected in a convenient manner that challenges the human participants to learn the rewards of the tasks without making the task overwhelming (e.g., if it is too noisy there is nothing to learn; if there is no drift at all then the task is too easy and after a few decisions there is nothing to learn).

[0054] Regarding the reaction times, this information is also recorded, but reaction time is not central. However, this does not mean that we will never use this information, as in the end reaction times can provide some information about the decision processes during a bandit task. Therefore, this information can also be used to improve prediction accuracy.

[0055] Further embodiments of the invention are laid down in the dependent claims.

[0056] BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Preferred embodiments of the invention are described in the following with reference to the drawings, which are for the purpose of illustrating the present preferred embodiments of the invention and not for the purpose of limiting the same. In the drawings,

[0058] Fig. 1 shows a sketch of a system of the invention;

[0059] Fig. 2 shows the step of one assessment of behaviour and pupil size with an application of the system of Fig. 1 ;

[0060] Fig. 3 shows the development of the reward in a two option embodiment of the system, being preferentially an answer-dependent development of the reward;

[0061] Fig. 4 shows the representation of the decision tree of the control system over a sequence of trials as shown in Fig. 3 according to an embodiment of the invention.

[0062] DESCRIPTION OF PREFERRED EMBODIMENTS ETH 2023-172

[0063] 8

[0064] Fig. 1 shows a sketch of a system 100 according to an embodiment of the invention. The system 100 comprises a control unit 10, usually a programmable computer system comprising different input and output units as components. The system further uses the following components:

[0065] (i) a sensor 20 measuring pupil diameter of one eye or of two eyes simultaneously. The sensor 20 can be part of a camera, especially of an infrared camera. It can be a stand alone camera or integrated into another device, as e.g. a smartphone or via camera contact lenses. The sensor 20 is mounted to an object in the environment or directly fixed to the body of the user 5 of the system.

[0066] (ii) preferably, an ambient light sensor / camera 30 measuring ambient light conditions mounted to an object in the environment or directly fixed to the user 5 of the system.

[0067] (iii) preferably, a luminance fluctuation compensation unit 40, e.g. a hardware solution that compensates for fluctuations in luminance of the environment around the sensor / camera 30. Potential hardware solutions include light adaptive glasses, goggles or contact lenses, or static / mobile virtual / augmented reality setups provided to and used by the user 5 of the system.

[0068] (iv) a sensory stimulation module 50 that provides visual, tactile / haptic, auditory, gustatory, or olfactory stimuli that the user 5 employs to interact with the system 100 via the performance of cognitive / decision-tasks of any sensory modality. The sensory stimulation module 50 might be integrated in the hardware component of sensors that acquire other (neuro)physiological signals such as ECG, (in-ear)EEG, etc. .

[0069] (v) a choice detection unit 60 configured to detect the response of the user 5. This can be an input on a computer keyboard or the detection of the fixation of a part of a screen related to the choice of the user 5 or any other corporal movement of the user 5 expressing their choice. One of the simpler choice detection units 60 is the use of a mouse with dedicated choice fields on a computer screen, especially when the stimulation from module 50 is a visual stimulation provided on the same computer screen having the choice fields. A keyboard or a keypad can also be used as choice detection unit 60

[0070] (vi) an analysis unit 70 that analyses the continuous temporal dynamics of pupil dilation as reaction on the stimuli of module 50. The analysis unit 70 comprises a piece of software incorporating a measurement algorithm to estimate presymptomatic stage progression of AD and brain age progression.

[0071] In the present context, the sensory stimulation module 50 presents a number of choices or selections to the user. The specification uses the term "selection" for the choice options left to the user to differentiate the term over the actual choices made by the user 5 and retrieved ETH 2023-172

[0072] 9 by the choice detection unit 60.

[0073] Preferably, the system 100 and apparatus is connected to a data centralization platform that manages, stores, and delivers information to authorized personnel / users.

[0074] Essential features of the application of the system 100 comprise:

[0075] - Assessment of pupil size (one eye or two eyes) over time with a certain temporal resolution, preferably light condition correction, and timelocking to sensory stimulation 50 or specific user behaviour as captured by the detection unit 60, i.e. the system tracks from the user the pupil reaction as well as the answer time on the challenge and sensory stimulation.

[0076] - Pupillometry in combination with situations that challenge the function of the arousal system (e.g., involving decision policy learning).

[0077] - A causal inference model was configured to directly use ongoing pupil dilation and decision-making dynamics as a requirement to characterize and predict brain age and presymptomatic AD (Figure 4). Adaptations of causal inference models can be implemented with any type of dynamical system (e.g. recurrent neuronal network, a transformer network) to handle dynamic data (e.g, that depends on the temporal structure and also may depend on history information (i.e., information before the current trial t, e.g., t-1) of input variables of interest like pupil dilation (Unit 402’), choice behavior (Unit 403’), and the state of the environment (Unit 413); any other variable relevant to the task at hand can also be used as input to the system such as additional physiological measurements of the user (e.g., EEG, ECG), user demographics, etc). The use of such a dynamic causal inference model allows exploiting the trial-by-trial dynamics of the user’s decisions / actions aton each trial t by conditioning the user’s decisions / actions on the potential condition c of the user, e.g., c=1 indicating Alzheimer’s present or c=0 Alzheimer’s absent, where dynamical Units 440’ and 440” predict the user’s decision / action at on each trial t given c= 1 (Unit 440’ in Figure 4) and c=0 (Unit 440” in Figure 4), respectively. Another example can be to use condition c as an indicator of age (c=7 old; c=0 young), but c could also be a continuous variable indicating the predicted age, e.g., in years. The advantage of using this example dynamic causal inference model configuration — that exploits the dynamics of the trial-by-trial user’s actions / decisions alongside pupil measurements — is that it also allows the incorporation in the same system of trial-by-trial propensity score inference TT(C) through an overarching dynamical system (e.g., an additional recurrent neuronal network, a transformer network; ETH 2023-172

[0078] 10

[0079] Unit 400 in Figure 4). The inferred propensity score TT(C) (Unit 430 in Figure 4) delivers — for each trial t — the likelihood of belonging to the Alzheimer’s disease group, i.e., the probability T(C= 1) (or the likelihood of belonging to the healthy group, i.e., the probability TT(C=0) = (1- TT(C= 1))). A cumulative sum of these log-likelihoods ( i.e., Iog[7r(c= 1) / (1-TT(C= 1)] ) across all trials t — that the user engages in — provides a summary score of the log odds that the user belongs to group c=1. It is an advantage of the system when the causal interference model can be fine-tuned / calibrated to unique data acquired from confirmed pre-symptomatic AD individuals who are cognitively unimpaired and whose cognitive symptomatology would escape diagnosis based on standardized assessments. Other configurations / architectures of such dynamical causal inference models are possible.

[0080] Fig. 2 shows the step of one assessment (defined here as a trial) of behavior and pupil size with an application of the system 100 of Fig. 1. Box 201 is related to option presentation by the sensory stimulation module 50. The option presentation 201 is shown to comprise a choice between two options A and B. It is possible that the sensory stimulation module 50 uses three or more options, as e.g. four options A, B, C and D (not shown). The option presentation 201 can be related to different modalities of task: this can be a visual information with an input relating to this decision. It can be a haptic input and decision as information transmitted to e.g. the underside of different fingers but likely also other modalities as sound based (high - low frequency or high and low intensity, other tactile, smell, taste etc information which can be provided as a discrete option of choice.

[0081] The type of task in Fig. 2 is a two-arm bandit task, i.e. decision-making between two choices with a hidden reward structure was used but other tasks that resemble similar scenarios (uncertainty, cognitive load, learning spatial or temporal or cognitive maps) could be used in other embodiments.

[0082] A time of e.g. 1000 milliseconds exists between the option presentation 201 of the sensory stimulation module 50 and the fixation step 202. The fixation step 202 is accompanied by a pupil reaction detection by the sensor 20.

[0083] The fixation step 202 is followed by the response choice and detection step 203. The response choice and detection step 203 comprise the choice of the user 5, here option A and the related input which can be detected by the choice detection unit 60.

[0084] After a predetermined time as e.g. 700 milliseconds, the control unit 10 presents the user ETH 2023-172

[0085] 11 the reward in a reward presentation 204. The reward is a point value, here the value 19. This concludes one trial step and is followed by the next options presentation 20T as start of the next trial.

[0086] It is the task, as explained before the start of the sequence of trials, of the user to obtain, after the number of trials, the maximum number of points, i.e. , the sum of all rewards of all trial steps as repetitions of the steps as shown in Fig. 2. It is possible that the user 5 knows or does not know the maximum number of points as reward in each trial. Of course "maximum number of points" is one option. The task can also be to obtain the least number of points or to strive for the least deviation from a specific level of points, as e.g. 50 etc. .

[0087] Fig. 3 shows the development of the reward in a two-option embodiment of the system. As explained below the system is preferentially an answer dependent development of the reward. The x-axis reflects the number of trials. The number of trials chosen can be 100 and announced before the start of the exercise to the user 5. The y-axis shows the values of the rewards which are between 0 and 100 in a two option system with the aim to obtain the highest (or e.g. lowest) number of points after the trials. The curve 301 shows the reward of option A and the curve 302 shows the reward of option B as a choice of the user 5.

[0088] What is not shown in Fig. 3 and embedded in the control unit 10 is the development of the reward between the two options. The development of the reward for each and between the two options can be predetermined with independent values, or the sum of all options can be e.g. 100, etc. The reward of the next option presentation 20T can also be calculated as response of the previous choice (= answer) to enable the system 100 to be obtain "better" answers by the user. The term "better" refers to a more challenging trial presentation to have more distinctive pupil reaction values within the chosen number of trials (here 100). This approach avoids a high number of trials which high number would not be easily accepted by the user 5. In other words, the "next reward" determination strives to avoid reaching a plateau where the user 5 is maintaining previous choices as answers or where the user is always alternating between the two options within his answer. In fact, any simple detectable pattern of reward determination is avoided. This is shown in Fig. 3 where no predictable pattern can be seen in the reward evolution.

[0089] The value of the hidden reward for each bandit or choice option (the left (bandit A) or the right bandit (bandit B) in Fig. 2) that is chosen in each trial is revealed to the participant via visual feedback using a numerical value displayed on a computer / tablet / mobile-phone ETH 2023-172

[0090] 12 screen. However, one can use any other sensory modality to reveal the value of the chosen bandit. For example, one could use auditory feedback. If there are N=2 bandits, and the human participant uses her left hand to choose bandit A (or right hand to use bandit B), then the hidden reward of the chosen bandit can be revealed with a voice that states the numerical value explicitly (e.g., “the chosen bandit has a value of 15 reward points”).

[0091] Fig. 4 shows the dynamic causal inference model (explained above) of the computer system 10 according to an embodiment of the invention. It is clarified that the general goal of the neural network 450 (right-hand side of panel a in Fig. 4) is to obtain a low-dimensional embedding stthat summarizes the pupil dynamics at around the time or the decision / action taking place at trial t. This embedding can be simple, such as a scalar value expressing mean pupil dilation at around trial t, or it can be set to capture a richer complexity of the pupil dynamics but still keeping it low-dimensional. This low-dimensional embedding can be derived using for instance a PCA, a neural network, or any other related method that provides a compressed yet meaningful representation of the neural (e.g., pupil) dynamics at around trial t.

[0092] The goal of the behavioral task is to induce meaningful pupil dilation dynamical changes over time that allow the computer system software to be sensitive in predicting aging and Alzheimer’s risk. Performing any task that challenges human participants by changing from learning / exploration stages to learned / exploitation states, and back, is considered to induce neurophysiological engagement of brain structures thought to be affected during the early stages of Alzheimer’s disease, e.g., the locus coeruleus (LC) and the orexin / hypocretin neurons (HON).

[0093] Thus, generally speaking, a detailed study of pupil dilation dynamics during decision-making tasks involving the learning of abstract or objective cognitive maps may not only allow the development of predictive markers but also gain insights into the neurobehavioral mechanistic processes underlying the presymptomatic stages of AD and biological brain aging.

[0094] The further input to the model includes the calculation of a propensity score for each trial that is compared to a calibration dataset. This comparison determines the log-likelihood of the trial data indicating a particular risk for condition / state.

[0095] Calibration data is available from individuals who are known to develop AD but who are still ETH 2023-172

[0096] 13 at the presymptomatic stages. This is possible, for instance, based on genetic mutation carriers with deterministic AD. Using such calibration data the functional validity of the invention is confirmed, where it is shown that it is possible to predict presymptomatic AD individuals, where the incorporation of pupil dynamics is fundamental to generate accurate predictions of the disease condition (Figure 4, panels b and c)). Additionally, such a system can learn based on retrospective datasets decades before AD is diagnosed later on. However, the employed method here allows for long-term assessments and future calibrations of the slope of brain aging and possible dementia onset, given the continual learning characteristics of the present method.

[0097] Such a dataset for calibration of Alzheimer's Disease risk can comprise:

[0098] - presymptomatic / asymptomatic individuals with genetic markers indicative of an increased risk or deterministic for Alzheimer's Disease

[0099] - presymptomatic individuals predisposed to Alzheimer's Disease who later develop symptoms (longitudinal data)

[0100] - control groups without risk factors for Alzheimer's Disease, and

[0101] - individuals with Alzheimer's Disease at various stages of disease progression.

[0102] ETH 2023-172

[0103] 14

[0104] LIST OF REFERENCE SIGNS

[0105] 5 user

[0106] 10 control unit

[0107] 20 sensor

[0108] 30 ambient light sensor

[0109] 40 luminance fluctuation compensation unit

[0110] 50 sensory stimulation module

[0111] 60 choice detection unit

[0112] 70 analysis unit

[0113] 100 system

[0114] 201 option presentation

[0115] 20T next option presentation

[0116] 202 fixation step

[0117] 203 response choice and detection

[0118] 301 reward for one option

[0119] 302 reward for one other option

[0120] 400 Dynamical system (e.g., recurrent neural network, transformer, etc) disentangling propensity scores and conditional decisions / actions

[0121] 400' Dynamical system (e.g., recurrent neural network, transformer, etc) predicting decisions / actions on trial / conditional on e.g., disease present: c=1

[0122] 400" Dynamical system (e.g., recurrent neural network, transformer, etc) predicting decisions / actions on trial / conditional on e.g., disease absent: c=0

[0123] 402' low-dimensional embedding of pupil dynamics at around trial /

[0124] 403' decision / action a of the previous trial t-1

[0125] 413 reward r experienced by the user in the previous trial t-1

[0126] 430 propensity score (i.e., probability of belonging to a group c)

[0127] 440' Probability of decision / action a in trial / conditional on the user belonging to group c=1 (e.g., disease present)

[0128] 440" Probability of decision / action a in trial / conditional on the user belonging to group c=0 (e.g., disease absent)

[0129] 450 System (e.g., neural network) computing a low-dimensional embedding of the pupil dynamics at around trial /

Claims

F0724715CLAIMS1. A pupillometry system (100) for computer-implemented assessment of brain aging and presymptomatic Alzheimer's risk in users through computer-human interaction with a user (5), comprising: a computer system (10); a pupillometry sensor unit (20) for at least one eye; a sensory stimulation unit (50) configured to present (201) the user (5) a selection of at least two discrete sensory information; a response choice unit (60) configured to detect (203) the choice of the user (5) between the discrete sensory information; wherein the computer system (10) is configured to execute a predetermined number of trials, wherein in each trial, upon receiving a response from the user on the presented choice, the computer unit (10) is configured to determine and present (204) a reward / outcome to the user (5) depending on their choice, wherein each trial comprises: the control sensory stimulation unit (50) is adapted to provide one predefined selection of sensory information (201); the pupillometry sensor unit (20) is configured to detect the reaction of the pupil(s) of the user (5) over time between the presentation of the predefined selections and the presentation of the reward / outcome and to transmit the pupillometry sensor information, especially pupil size over time to the computer system (10); the response choice unit (60) is adapted to detect the choice of the user (5) between the discrete sensory information as an answer and to transmit the answer to the computer system (10); and the computer system (10) is adapted to present the reward / outcome to the user (5); wherein the computer system (10) is adapted to calculate upon pupillometry analysis of each trial for the predetermined number of trials and for the entire trial series the probability of the AD presymptomatic risk through applying an inference model using as input vectors the reaction of the pupil(s) of the user (5) over time between the presentation of the predefined selections and the presentation of the reward as well as at least, where available, one of the previous answers on previous trials.

2. The system (100) according to claim 1 , wherein the sensory stimulation selection is from the group encompassing visual, acoustic, tactile, olfactory and gustatory stimulation.F07247163. The system (100) according to claim 1 or 2, wherein the sensory stimulation selection comprises two or more discrete selections.

4. The system (100) according to any one of claims 1 to 3, wherein the pupillometry sensor unit (20) comprises an ambient light sensor (30) and / or a luminance fluctuation module (40).

5. The system (100) according to claims 1 to 4, wherein the rewards / outcomes for the different trials follow a generative (pseudo-)random walk path during a trial sequence.

6. The system (100) according to any one of claims 1 to 5, wherein the presented selections are based on a bandit / cognitive-map system.

7. The system (100) according to any one of claims 1 to 6, wherein the pupillometry analysis of each trial for the predetermined number of trials and for the entire trial series comprises the comparison with a predetermined database of pupillometry analysis data of persons from the group comprising asymptomatic individuals with markers (e.g., genetic) indicative of increased risk or deterministic for Alzheimer's Disease, asymptomatic individuals who later develop Alzheimer’s disease (based on longitudinal, prospective, or retrospective data), control groups without risk factors for Alzheimer's Disease, and individuals with Alzheimer's Disease at various stages of disease progression.

Images