WO2021083512A1 - Measuring an attentional state and providing automatic feedback during a technical system interaction - Google Patents

Abstract

The invention refers to an approach to improve a user's attentional skills and to stay focused. In particular, the invention relates to a method for providing a feedback signal during a user's control session, comprising interactions with an application, based on measured signals. The method comprises to determine the application to be monitored and to measure input signals (in), comprising a set of interaction signals from at least a group of sensors, which track a user's interaction behavior during the control session. The method further uses a computing unit (100) for continuously inferring (S3) attentional states (as) over time within the control session, from the measured input signals (in); and for calculating (S4) a formative feedback signal (fbf), based on the inferred attentional states (as) to be provided on an output interface (102) in case a change of the attentional states (as) has been detected; Further a stored model of plasticity of cognitive control (LVOC) is accessed for calculating (S6) an estimated control signal (pred-ct+1) and an estimated attentional control skill level (pred-skt+1) by using a Bayesian inference algorithm based on a prior probability distribution of the attentional control skill level and the control signal; an improvement signal as summative feedback signal (fbs) on an output interface (202) in case an improvement of the estimated attentional control skill level has been detected with a pre-configured confidence value.

Description

Title: MEASURING AN ATTENTIONAL STATE AND PROVIDING AUTOMATIC FEEDBACK DURING A TECHNICAL SYSTEM INTERACTION

The present invention relates to a computer system and a computer-implemented method (e.g. an app) for providing automatically calculated feedback during a technical system interaction, e.g. a user interaction for controlling an operating method for a technical device (e.g. an automation system, a medical diagnosis apparatus or a robotic application in a production system), and thereby improving control capabilities.

In a variety of technical fields, modern systems are getting more and more digitalized, e.g. in the domain of medical engineering, robotics, or control of other complex technical systems involving a high degree of responsibility and concentration. That is to say, the control of a technical system is done by means of at least one computer application which is operated and controlled via a human-technology-interface (e.g. user interface).

Usually, on a computer system a variety of different applications are installed and may be operated. Thus, these “other” applications may distract user’s attention for his very control task. Further, other context or environmental conditions may also distract the user's attentional skill. This, in turn, several settings, parameters and circumstances may negatively influence the technical system to be controlled and may lead to errors and/or failures.

The problem of being distracted is therefore an important issue in the context of controlling a technical system.

It is therefore an object of this application to provide a mechanism to improve a control skill level of a user or agent during his interaction with a computer application. In particular, the present invention refers to subject matter according to the appended independent claims.

According to a first aspect the present invention refers to a computer-implemented method for providing a feedback signal during an agent’s control session, comprising interactions with at least one application. The application may be an application for controlling a technical system (e.g. in the field of medical informatics or production or automation). The method is based on measured signals, wherein the method is executed on a computer system. The method comprises:

Determining at least one technical application to be monitored; (there may be one or more technical application which is to be tracked); - Measuring input signals, comprising a set of interaction signals from at least one sensor of a group of (technical) sensors, embedded in a Human-Technology- Interface (like a Human-Machine-Interfaces or a Human-Computer-Interface) which detect and track the agent’s interaction behavior during his control session;

- Computationally inferring attentional states continuously over time during the control session from the measured input signals;

- Using a first computing unit for calculating a formative feedback signal, based on the inferred attentional states to be provided on an output interface in case a change of the attentional state has been detected (with other words: if the actual attentional state is deviating from the prior attentional state);

- Accessing a storage with a stored model of plasticity of cognitive control and providing the inferred attentional states, a control signal and the provided formative feedback signals and session parameters to the model for calculating an estimated control signal and an estimated attentional control skill level by using a Bayesian inference algorithm based on a prior probability distribution of the attentional control skill level and the control signal;

- Using a second computing unit for providing an improvement signal as summative feedback signal on an output interface in case an improvement of the estimated attentional control skill level has been detected with a pre-configured confidence value. The first and the second computing unit may be the same or may be integrated in a common superior computing entity.

The agent may be a computing agent (robot) or a human agent. The following description is related to a ‘user’. In this context, the term has to be construed to comprise a human user and a robotic, artificial, or electronic controlled user. In the latter case, a humanoid robot interacts with the control application via an input interface.

In the ‘human user’ case, the background of this solution is based on the finding that many people (e.g. students) struggle to stay focused. Also, students have problems to stay focused for long enough to learn effectively and the rise of social media is making this problem still worse. Constant distractions at work cost the economy billions of dollars’ worth of losses in productivity. These serious issues harm not only the life of individual people but also society as a whole. In this application, the challenge of staying focused in the face of distractions is addressed by providing a brain training application (in short app) called ACTrain. ACTrain will be a personal assistant with a name and a custom appearance that trains people to stay focused on a task and effectively resume it after getting distracted. Unlike conventional brain training apps, ACT rain allows people to train while they are working or studying, thereby turning their daily life into a gym for their mind. This makes it possible to use ACT rain in many different contexts including technical applications and workplaces as well as in education. The heart of ACT rain is an intelligent feedback mechanism based on computational models of how attention control skills are learned. Based on these models, the application gives people feedback when they get distracted. The feedback communicates the benefits of regaining focus for their productivity and success. The provided software may be used within the context of e-learning and online courses. The software will benefit the life of millions of students and working professionals.

Thus, in one preferred embodiment, the invention refers to training the attention control skills of knowledge workers (e.g., while they are writing a text e.g. in Word or working with an e- learning app). In another embodiment the method and apparatus may be used in the context of therapy, in particular for an ADHD therapy. For this purpose and according to an embodiment, keyboard and mouse inputs are used for model-based calculation of a feedback and to provide this feedback to the person while he or she writes a text or executes a task on the computer.

Accordingly, in a preferred embodiment, the control session is a learning skill task and wherein an adaption signal is provided for dynamically adapting a difficulty level of the task and/or a duration of the control session, based on the formative and/or summative feedback signal.

In another preferred embodiment, the input signals do not only comprise the interaction signals but further comprise a plurality of different other signals. The multi-modal signals may be used for verification of the detected signals. For example, if four different signal sets show similar characteristics over time and another fifth signal significantly deviates from this, then, it is inferred, that the fifth sensor might be misplaced or the fifth signal is a failure/artifact. Generally, the other signals may serve to measure neurophysiological and/or peripherphysiological activity from a set of multi-modal neurophysiological and/or peripherphysiological sensors that can be attached directly to the human body or can be integrated into textile fibers, clothes, and/or elastic bands, to provide continuous signals. These signals may be processed in order to estimate an attentional state of the user and of gradual changes of the user’s attentional state over time and/or momentary attentional state.

In a preferred embodiment, at least a first signal being an peripherphysiological signal, in particular an eye tracking signal (and inferred eye features) and another second signal, being a neurophysiological signal, in particular e.g. an EEG signal, is measured and processed for calculating or inferring the attentional states. In a preferred embodiment, the first and second signal are co-registered or co-related. The co-registration procedure will be explained in more detail below.

The method is computer-implemented and is executed by means of a feedback application (which is not the application to be monitored, but refers to the computer program which executes the method, described above).

In a preferred embodiment, the feedback application can be divided in four major parts:

1. A set of triggers

2. A tracker

3. The user interface

4. An orchestrating object.

In particular, a family of small tasks, each running in a separate thread, that monitor some aspect of user activity (so called _triggers_), an object, that consumes the signals emitted by the triggers and decides whether the user is performing intended activity or are they distracted (so called _tracker_), the user interface, and the orchestrating object that synchronizes, reacts to and controls all of the units, that comprise the application (this object is tautologically named as _application_).

Each of the triggers operates in a separate thread and use an asynchronous message queue to communicate to the rest of the application. In a basic embodiment, there are three _triggers_:

1. The keyboard and the mouse listener are trying to determine whether the user is performing any activity at all.

2. Active window title monitor periodically queries the operating system for the title of the currently focused window.

3. Periodic reminder which emits a signal every n-seconds when a user is determined to be distracted.

To decide whether a user is performing a desired activity (task) or not, the _tracker_ is listening for the signals emitted by the triggers. The logic is trivial: if any input has been lacking for a significant number of time, the user is assumed to be distracted; if the active window cannot be associated to the desired activity/task the user is considered distracted as well. This window classification is done based mostly on the window title and the application name and is performed by a separate entity called _window classifier_.

The only information about the currently focused window we can reliable get from the operating system across all the platforms is the window title. On MacOS and Linux it is also possible to query for the human-readable name of the application that the window belongs to. On Windows we can only get the base name of the executable file itself. There are very few ways one can classify windows as fitting to a certain activity based on this information alone.

A first option is to use a list of keywords one might expect to find in a window title. For example, if the user is preparing a presentation for a talk, we could expect to see "power point", "slides", "keynote" or even "Presentation" in the title of the presentation program. Thus, any window that has either of those keywords in the title will be considered to be a window of a presentation program.

This approach is not very reliable. One problem that we have encountered were the dialog windows and the menus, especially on Windows platform. To illustrate suppose that the user is preparing on a presentation. They open a new power point window, and the window title says "POWERPNT.EXE Prasentationl - PowerPoint". Since the keyword "Presentation" is in the title, the window is classified as a part of the presentation software. However, as soon as they decide to switch the default font and they open a fonts menu, the title of the currently focused window (at least from the point of view of the operating system) switches to "POWERPNT.EXE Schriftart". None of the keywords are seen in the title and this window would be immediately classified as not being a part of the presentation software.

Therefore, (to work around that) a status cache is used, based on unique process id. We use a dictionary that maps the unique process id (PID) of the application to the window status (either fitting to the activity or not). When we want to classify the window we take the window title, the application or the executable name and the PID. We first do a cache lookup. If we find the current PID in the cache _and_ the application is considered to be allowed then we return the cached value. Otherwise, either we cannot find the PID in the cache, or the application is marked as not allowed, we classify the window based on the window title and store result in the cache. Thus, if the application was considered appropriate for the activity once during the training session it will be considered as such until the end of the session.

Thus, preferably, the set of input signals and in particular interaction signals which are used for controlling the application may comprise a window activation signal, indicating whether the window is currently used and to which application it belongs. As explained above, the window activation signal may be generated by caching a window classifier from an operating system. The window classifier is a unique process-identifier (PID). It may be read out from the active application on the operating system. It may be stored in a dictionary or another field in a storage unit. The window classifier represents, which application is active and actively used. The window classifier may comprise an entry in the z-order list, defined by the operating system to display the different application windows in a virtual z-order on the screen, thereby representing which windows are currently used for interaction (e.g. pop up windows and message windows are always displayed at the topmost position). For example, if the requested input in such a popup window is provided by the user, the system may automatically infer that the user’s attention is focused on the very window and thus on the determined application to be monitored.

In another preferred embodiment, the neurophysiological and/or peripherphysiological activity may be measured in reply to a sequence of formative and/or summative feedback signals provided by the computing unit. The feedback signal may be provided on an output interface as well, e.g. in the form of a visual component, an auditory component and/or reward/distraction message.

In a preferred embodiment the neurophysiological activity is measured by means of a functional near-infrared signal (fNIRS), representing human executive functions in predetermined prefrontal cortical brain regions.

According to still another embodiment, the neurophysiological activity is measured by means of an electroencephalogram (BEG) signal or categories thereof (like e.g. oscillatory-based frequency measures), comprising event-related potentials, oscillatory-based frequency measures, time-series and frequency-based connectivity measures, related to user’s attention and/or perceptual user feedback processes.

The peripherphysiological activity may be measured by means of an eye-tracking and/or pupillometry signal, which is used by the processing unit to estimate eye features such as fixation duration and pupil dilation activity related to user’s attention and/or perceptual user feedback processes.

The measured eye features, in particular a fixation onset, may be used to estimate and corelate fixation-related categories of an BEG signal, such as fixation-related potentials related to higher order cognitive processes such as stimulus perception (e.g. feedback-recognition) and allocation of attentional resources for controlling the computer application.

The peripherphysiological activity may also be measured by means of an electro-dermal signal and/or heart-rate signal, related to user’s arousal and/or perceptual user feedback processes.

In a preferred embodiment, a set of optimal features or categories from the neurophysiological and/or peripherphysiological signals adaptively optimizes inferring the attentional state of the user in a way that minimizes a prediction error by selecting the sequence of optimal feedback signals using a Bayesian inference algorithm. Thus, in a preferred embodiment a configuration menu is provided in a preparation or configuration phase of the app. The configuration menu comprises a set of selection menus, input fields or buttons for selecting the type of input signal to be used for inferring the attentional states. This has its background in the fact that for specific use cases and type of tasks (e.g. control task for which kind of system, reading task, writing etc.) a corresponding specific and potentially limited set of input signals are relevant for inferring the attentional state. Further, the configuration menu may comprise menus, input fields and/or buttons to select and/or determine a specific category of the respective input signal. For example, if the heart rate has been selected as input signal, then, different components of the signal may be determined, e.g. based on the spectral analysis of the signal, in particular a high vs. a low-frequency spectrum and/or a heart rate variability representing different modulations of the autonomic nervous system. Thus, different components, parts, ranges and/or spectrum of the signal may be determined for being processed.

The computationally inferred attentional states (which are inferred by an algorithm from the input signals) and optionally together with the corresponding input signals may be stored in a user-specific and/or task-specific context profile. The context profile may be used to manage and distinguish relatively stable attentional user traits, and/or gradual changes over time and/or momentary attentional user states.

In a preferred embodiment, data synchronization methods, may be applied to synchronize the different types of input data from the distributed sensors and entities. Synchronization methods may for example be a lab streaming layer (LSL), which is a framework allowing time- synchronized recording of data streams from multiple modalities, such as EEG, eye tracking etc. Preferably the synchronization methods are used for correlation and co-registration.

In a preferred embodiment, the formative feedback signal is provided for several times during execution of the control session but only in case a change of the attentional state has been detected, which means a change between the actual state and previous states. This has the advantage, that the user directly gets feedback on his attentional skills during task execution and is reminded to further stay focused and not to get distracted.

In another preferred embodiment, the summative feedback signal is provided after each execution of the control session or after a pre-defined set of control sessions. This has the advantage that the user will not be disturbed by the summative feedback signal. In particular, if the summative feedback signal is provided too often, the effect on improving the attentional control skill level will be reduced. The determination of the pre-defined set of control sessions may be based on rules, e.g. “summative feedback signal is only to be provided in case the confidence level is over 90%, and more preferred over 95% or if the improvement is positive".

Estimating an attentional control skill level may comprise calculating a value function of a Markov Decision Process model of skill acquisition.

In another preferred embodiment, a guidance-fading-out-mode is applied or used which adaptively controls a frequency and/or intensity (e.g., brightness, signal strength) for providing the formative feedback signal. Optionally, the guidance-fading-out-mode may also be applied for providing the summative feedback signal. The guidance-fading-out-mode may be adjusted in response to pre-configured rules. The rules may be stored e.g. in a rule database and /or on a central storage entity. This has the technical advantage that the app may be better adjusted to the individual needs of the user. Thus, the provided support better aligns to the detected skill level (e.g. a higher skill level or attentional improvement only requires the feedback signal to be presented less frequent).

Preferably, the processed signals and/or data, in particular the measured interaction signals, the measured set of multi-modal neurophysiological and peripherphysiological signals, the inferred attentional states, the formative feedback signal and the summative feedback signal are stored locally. For improving security in another preferred embodiment, an encrypted data package is generated from the processed signals and/or data for preparing the encrypted data package to be securely sent to a server via a network connection. For example, by using a TCP/IP protocol a local archive may be generated for the encrypted data package, which increases privacy of the data. The server is in data connection with the model and may access and amend the model.

A further embodiment is directed to securely transferring the experimental data (measured and calculated signals) to a central server, which may be operated by researchers. All the user activity during the training session is written in a session-specific file, together with the current configuration of the application, and the user id. All the data is anonymized, and the data are provided to be non-accessible by anyone except the researchers. To achieve that the following encryption protocol, in particular asymmetric encryption protocol, is employed:

1. A pair of RSA keys are generated in advance. The private key is kept by the researcher / server responsible for the experimental data processing and only by them. The public key is put in the code repository and built into the binary executable file that 1 \ distributed to the users.

2. When the user is done participating in the experiment and want to submit the session data to the server, they press the "Create submission package" button. The application then finds all the session files and compresses them into a single archive.

3. A random 256-bit encryption key is generated. It is used to encrypt the archive using symmetric AES encryption algorithm. The randomly generated key is then asymmetrically encrypted using public RSA key shipped with the application. This encrypted key is attached to the end of the encrypted archive. This encrypted package is saved on a disk and the dialog with the file location is shown to the user.

4. User manually submits the encrypted package to the server, either through a submission form or via an email.

5. To open the submission package the researcher (server side) uses a special application that requires the private RSA key. First the encrypted AES key is detached from the package and private RSA key is used to decrypt the key. Then, using the AES key the package is decrypted using AES algorithm. Finally, the resulting archive is deflated.

Thus, to sum it up, in this embodiment a symmetric AES encryption with a random 256-bit key is used to encrypt the package and an asymmetric RSA encryption is used to encrypt the AES key. Both, the encrypted archive and the encrypted key are transferred to the researcher’s server. On their end the AES key is decrypted and then the key is used to decrypt the archive. Since the AES key is randomly generated and then asymmetrically encrypted using public RSA key, it is impossible to decrypt the package without the private RSA key, that is supposed to be only in the hand of the researchers. This embodiment has the technical advantage that security is improved and compromising the transferred data package between local user and central server may be prevented.

In another preferred embodiment, the method comprises to support refocusing after distractions by providing a resumption mechanism based on the (recorded) set of signals. This is an important advantage and may further improve the attentional skill level. The resumption mechanism may comprise to control the output of the application in an adapted manner such as the last window will be presented in a highlighted manner. Alternatively, the last fixated area before getting distracted (by means of the eye tracking signal) on the screen may be highlighted. Alternatively, an acoustic signal may be provided with a reminder and a relation to the last window or application which was used.

Preferably, estimating an attentional control skill level is executed by using a Bayesian inference algorithm. The Bayesian inference algorithm is preferably applied to all the signals, in particular to the neurophysiological and/or peripherphysiological signal set and the predictions of a model of the plasticity of cognitive control are used to define a prior probability distribution over the user’s attention control skills in the next time step.

The formative and/or summative feedback signal is/are preferably provided on an output interface, in particular on a user interface. Each of the feedback signals may comprise a message in textual form and/or a graphical representation and/or may be provided as sound signal and/or as control signal for controlling priority of the applications running in parallel with respect to the determined application, in particular a control signal for a z-order of a graphical user interface of the computer system.

In a preferred embodiment, a Kalman filter or another tool may be used for the set of multimodal neurophysiological and/or peripherphysiological sensors signals and/or for the interaction signals in order to improve the result in case of an uncertainty of the detected signals. Generally, a Kalman filter may be used for processing the detected or measured signals and for providing a prediction for further signals and to continuously check the provided predictions based on the actual measurements (in a loop structure). In general, the attentional state cannot be determined, or predicted exactly due to measurement noise, random inputs to the system or incomplete knowledge of the system parameters. However, although the state cannot be precisely determined, it is possible to estimate it in some optimal fashion by the appropriate processing of available data. Thus, the problem is that of estimating the state of- a system based on the available measurements and on knowledge about system parameters and noise statistics. Given the input measurements and some statistical properties of the noise and the past- history of input measurements, the question to be answered is, what is the best estimate or prediction of the state at some given time, in particular in future. The Kalman or related filters such as least squares, maximum likelihood of Bayesian estimators, may be used for state estimation of a dynamic system. All of them reduce to the Kalman form under assumptions of Gaussianess of random sequences and first order Markovian properties. Thus, briefly, the attentional state of to be determined is processed as a vector quantity which encodes all of the system history that needs to be known for predicting attentional skill level properties. Although there are many published applications of Kalman Filters, they have not yet been applied to attentional control tracking systems. Kalman filters benefit from the information about the attentional and learning skills for enhancing the accuracy of the estimation of future skills. Generally, the Kalman filter may comprise two stages: A prediction stage and an update stage. Initially, a prediction of the current attentional control state is made using the previous measurements and the precious state. This estimate is updated using the observations weighted by a certain function within the Kalman algorithm. For certain applications of the Kalman filter many operations must be performed on matrices as rapidly as possible. Thus, the computer for running the application may be a small or thin client but is a specially organized, general purpose computer which is particularly efficient for performing matrix operations. Incorporating matrix operations into the structure of a computer suggests the use of arrays of logic elements.

For combining uncertainty of more than one measurement signal, preferably, the system will invert a probabilistic generative model of how the various measurement signals are generated from the user’s (latent) attentional state by analytically calculating the posterior probability of attentional state given the measured signals according to Bayes theorem or using an approximate Bayesian inference algorithm (e.g., variational inference or MCMC; see Chapter 1.2 in Bishop, C. M. (2006). Patern Recognition and Machine Learning. New York: Springer). Further, a hybrid adaptive brain-computer interface, which his implemented as a generic Bayesian model, described as Markov decision process (MDP), which is a discrete time stochastic control process, wherein the prediction of the user’s attentional state is adaptively learned during the control session and wherein the inference of the user’s attentional state is optimized by minimizing the prediction error with the optimal selected feedback signals. Thus, there is a twofold feedback or loop between the input signals and the feedback signals.

In another aspect the invention relates to using the computer-implemented method as described above or a computer-implemented feedback module for the treatment of attentional disorders selected from the group consisting of an attention deficit syndrome (ADHD) and other related diseases with a malfunction of a user’s attention. Alternatively, the computer- implemented method may also be used for improving the attentional skill of a healthy human (e.g. for knowledge workers during reading and/or writing text) and help them to stay focused on their work. Alternatively, the method may also be used for robotic agents in an automation system. In the following a module is described, which is adapted to execute the method for providing a feedback signal as already described above. Features, advantages or alternative embodiments which have been described or claimed with respect to the method may also be applied and used to the module or system or computer program and vice versa. In other words, claims for the module or systems can be improved with features described or claimed in the context of the method. In this case, the functional features of the method are embodied by structural units of the system and vice versa, respectively.

In another aspect the invention refers to a computer-implemented feedback module for providing a feedback signal during an agent’s (e.g. a user’s) control session, comprising interactions with at least one application, based on sensor signal measurements, comprising:

- A determination unit for determining the at least one application to be monitored;

- An input interface which is adapted for measuring a set of input signals, comprising interaction signals from at least a group of technical sensors, embedded in a Human-Technology-Interface which are adapted for tracking the agent’s interaction behavior during the control session;

- A processing unit for continuously inferring attentional states, related to the control session, in response to the measured interaction signals;

- wherein the processing unit is further adapted for calculating and providing a formative feedback signal, based on the inferred attentional state on an output interface in case a change of the attentional state has been detected; and

- wherein the processing unit is adapted for accessing a model of the plasticity of cognitive control for estimating an attentional control skill level by Bayesian inference with a prior probability distribution, based on predictions of the model and a likelihood of the predicted feedback signals given potential attentional states; and

- wherein the processing unit is adapted for providing an improvement signal as summative feedback signal on the output interface in case an improvement of the estimated attentional control skill level has been detected with a pre-configured confidence value;

- The output interface for providing the formative and summative feedback signal in case a signal change has been detected.

According to a preferred embodiment of the computer-implemented module, mentioned above, the module is used for human agents and in addition comprises: - A set of physiological interfaces which are adapted for receiving a set of multi-modal physiological signals from a set of multi-modal neurophysiological and/or peripherphysiological sensors, comprising brain signals, indicating brain activity patterns during the control session. The brain activity patterns are indicative of an attentional state of a user.

In another aspect, the invention refers to a computer program comprising program elements which induce a computer system to carry out the steps of the computer-implemented method for providing a feedback signal during a control session of an application as described before, when the program elements are loaded into a memory of the computer system.

In still a further aspect of the present invention, a computer-readable medium is provided on which program elements are stored that can be read and executed by a computer system, in order to perform steps of the method for providing an output signal as feedback during a control session of an application as mentioned before, when the program elements are executed by the computer system.

In the following a definition of terms used within this application is given.

The “agent” may be an artificial agent, a robot or specific type of bot which is controlled by computer instructions. Preferably, the agent is a human user, interacting with a technical system via an application’s user interface.

The term “control session” refers to a session in which a user operates a computing device with at least one application for performing a certain task. The task may relate to a learning task or to a task for operating a technical system (actions in a production system, medical system etc.). In case of a learning task, several control sessions may be executed in sequence (e.g. a vocabulary learning app). The suggestion presented herein may also track the improvement of the attentional level over a sequence of such control sessions. The sequence may be consecutive or may be interrupted, e.g. by breaks.

The application is a computer application and may require user or agent input and user or agent control by means of a human technology interface, including Human-Machine-Interfaces and/or Human-Computer-Interfaces. The application relates to the task. If the task is a learning task, then the application is a learning application, e.g. an e-learning session. If the task is a control task for controlling a technical system (e.g. robotic system or an energy or production system), then the application serves for controlling the technical system. In certain use cases more than one application is necessary for controlling the technical system (task execution). Then, all the applications are taken into account when assessing the attentional control skill level of the user and the attentional skill level is evaluated for the complete set of applications, whereby a reference between the input signals and the respective application is processed and stored. Further, a weighting algorithm for the different signals may be applied in order to prioritize the signals in relation to their impact on the attentional control skill level and attentional state. Sensitivity and diagnosticity are considered for determining the effect of the respective signal on the overall impression.

The feedback signal may be any type of output signal, which has been automatically calculated by using complex calculations and by accessing a trained model. The feedback signal may be provided in different formats, e.g. in the form of a message to be presented on a user interface and/or as voice message via loudspeakers.

The provided feedback signal has two categories or types. A first type of signal is a formative feedback signal and a second type of signal is a summative feedback signal. The formative feedback signal is valid for the respective control session, i.e. during the control session. The formative feedback signal comprises the computed attentional state of the user while he or she is operating the computer for executing the application. The summative feedback signal is valid for a set of control sessions (e.g. several consecutive or sequential control sessions within a predefined period of time). The summative feedback signal is a more abstract signal than the formative feedback and represents the learning skill improvement or indirectly the level of success of the attentional level of the user over a particular time phase or over a set of control sessions. With other words, the summative feedback signal represents the attentional skill level learning or improvement process over time. An important feature of the present invention is that two different feedback signals are provided (formative and summative) in a different manner (with respect to time and repletion/signal strength). This has the technical effect that a metalevel for cognitive control may be provided and applied, which in turn improves efficiency of execution of the respective task (and in case of a control of a production system: efficiency of system control).

On the one hand the feedback signal is provided on a Human-Technology-Interface (in short HTI), e.g. a (graphical) user interface for the user and on the other hand the feedback signal is provided as input for the model for adaptively improving the model itself. In a first embodiment, the two different types of feedback signals are prepared to be shown at the same time and/or in parallel. In a second embodiment, the two types of feedback signals are prepared for being displayed separately and in particular independently, like e.g. in different time phases. Thus, the formative feedback signal may be displayed continuously in case a change of the attentional state has been detected, whereas the summative feedback signal may only be displayed or otherwise output after the control session has ended and/or if certain pre-configurable output conditions are fulfilled. An output condition for the summative feedback signal may for example be “a change of an average attentional state over a pre-configurable set of control sessions has been detected”. Other output conditions may be configurable. An advantage which improves flexibility of the application is to be seen in that the output conditions for the summative feedback signal and the formative feedback signal may be defined independently and thus may be different from each other.

“Measured signals” relate to the input signals, like neurophysiological and physiological signals or any other bio-signals which are measured during the course of a control session. The signals may be provided by sensors.

“Determining” the at least one application to be monitored may be executed by means of a selection menu with pre-configured items for user selection. Alternatively, a free text input field is provided on a user interface so that the user may determine the app he or she wants to track/monitor with respect to attentional behavior.

In a preferred embodiment physiological activity is measured by means of physiological sensors, which may include fixation-based sensors. The method thus comprises determining fixation-related signals. The sensors may be of different type and are adapted to measure different signals. Thus, the method may comprise a multi-modal neurophysiological data analysis for classification of mental states. Traditionally, EEG fixation related potentials (FRPs) have been the main tool of co-registration research reflecting higher order cognitive processing of stimulus information during reading and visual search applications. There are only very few studies investigating other potentially valuable fixation-locked sources of information beside FRPs. Accordingly, in another preferred embodiment, new markers may be applied and used for the calculations. Potentially new markers would be either fixation-locked and/or saccadic- locked oscillatory activity and connectivity measures estimated form the EEG signal. By using this co-registration approach (based on multi-modal signals, e.g. eye-fixation and FRP/EEG) we expect these measures to contain valuable information on how environmental conditions modulate cognitive and emotional (affective) processes. The purpose is to track stimulus perception (e.g. target-recognition), allocation of attentional resources (modality-specific workload states) and affective processing (valence-arousal) more precisely during real-world interactions using such multimodal measures, that are viewed from the onset of fixations as generated purposely by the human user. This approach would also provide a single-trial detection performance in these diverse scenarios, which could benefit a range of applications. From a basic research perspective, it allows us to study the dynamic interplay between attention and emotion (or affect) that is essential to better understand what drives human behavior while interacting during unconstrained situations in real-world applications.

To date, only fixation locked functional magnetic resonance imaging analysis for reading and language processing have been conducted, while other measures of hemodynamic concentration changes, as estimated via functional near-infrared spectroscopy (fNIRS), are still missing. Oxy- and deoxy-hemoglobin estimated from fNIRS provide an excellent tool for estimating mental workload and emotional experience, while being mostly invariant to motion artefacts. The combination of fNIRS with BEG is expected to increase brain state classification. In a preferred embodiment, fNIRS and BEG sensors are used for data acquisition and further processing.

While higher order FRPs research mainly focusses on attentional processes, the co- registration approach for studying emotional processes is important for providing the feedback signal. The term “co-registration" in this respect means taking the fixation onset (from eyetracking) as an event to analyze all the other measurement techniques from this onset. With this co-registration process it is possible to better capture the acceptance level of persons towards self-initiated autonomous and adaptive system behavior and may help to develop neuro-adaptive technologies that provide personalized assistance. In addition, even for applications that focus on attention, information on affective processes as acquired through measures that are especially sensitive to those processes, are expected to be helpful. In particular, the fixation-related potentials can inform about higher order information processing of stimulus information during reading and other tasks during the control session.

However, in other embodiments it is possible to define a different set of interaction signals to be acquired. For example, it is possible to collect BEG data in combination with pupil dilation signals. Preferably, the method may comprise a preparation phase for configuring settings of the method and an execution phase for executing the method for providing a feedback signal. In the preparation phase, e.g. the set of interaction signals to be used for later processing may be defined by means of buttons being provided on a user interface. The interaction signals are measured by means of sensors. The sensors may be embedded in a Human-Technology- Interface. Other sensors comprise neurophysiological and physiological sensors, BEG sensors, skin response sensors, heartrate sensors etc. The list is not limited to a specific kind of sensor. The sensors are adapted to measure analog or digital signals from a user while operating the application. Thus, with the sensor signals the user’s interaction behavior may be detected, e.g. when which kind of buttons are used and in which manner.

The measured interaction signals, as defined above, are then used by a computing instance (e.g. an algorithm) to compute or infer the user’s or agent’s current attentional state, while they are operating the application within the control session. Preferably, the measured interaction signals as well as the computed attentional states are stored in a specific storage and may be accessed for later processing steps.

After the formative feedback signal has been calculated, it preferably will be provided after the user has ended to execute the application, i.e. after the end of the control session. After a set of control sessions, it is possible to calculate and estimate the attentional control skill level by using a computing unit. The attentional control skill level is calculated by using or applying a Bayesian inference algorithm with a prior probability distribution, which is based on predictions of an attentional control skill level. The predictions are derived from a model of the plasticity of cognitive control which is stored in a memory. The attentional control skill level represents the likelihood of the attentional control skill of the user given the measured signals and computed attentional states.

The output interface of the computing unit may be a monitor or a set of monitor or windows to be presented of the monitor for providing formative feedback signal, a summative feedback signal and/or an improvement signal in case an improvement of the estimated attentional control skill level has been detected with a pre-configured confidence value.

The “model of the plasticity of cognitive control” is a computer-implemented model, which may be provided in a form which allows for adaptive improvements, based on the measured and provided signals. The model may for example be a model of learned value of control (LVOC), based on reinforcement learning. For further details it is referred to Lieder, F., Shenhav, A., Musslick, S., & Griffiths, T. L (2018) Rational metareasoning and the plasticity of cognitive control. PLoS computational biology, 14(4), e1006043.

A Bayesian inference algorithm is used for estimating the attention control signal that the user exerted from the measured signals and the LVOC model is applied to the estimate (or the corresponding posterior distribution of the attention control signal given the measurements) to predict the subsequent attentional control skill level (in future). The Bayesian inference algorithm is a method of statistical inference in which Bayes' theorem is used to compute the conditional probability that the user’s attention control signal assumed each of its possible values given that the measured signals were observed and an empirically-grounded prior probability distribution that encodes how frequently different attention control signals tend to be employed (for more information about Bayesian inference see Chapter 1.2 in Bishop, C. M. (2006). Pattern Recognition and Machine Learning. New York: brG^bή. In the following the invention is explained in more detail with respect to the figures.

These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a schematic representation of a method for providing feedback signals according to a preferred embodiment of the present invention;

Fig. 2 is another schematic representation of a method for providing feedback signals according to a preferred embodiment of the present invention;

Fig. 3 shows a feedback module, which may be implemented as “learning app" in an overview manner;

Fig. 4 is a sequence diagram showing the message transfer between different entities;

Fig. 5 is a flow chart of a method for providing feedback signals according to a preferred embodiment of the present invention;

Fig. 6 is more abstracted generalized representation of a learning app according to a preferred embodiment of the present invention; and

FIG. 7 is an overview figure for the architecture of the method according to a preferred embodiment of the invention.

DETAILED DESCRIPTION OF THE DRAWINGS The present invention relates to mechanism to improve an agent’s attentional skill level during executing a control session at a computer. A set of applications is implemented and executed on the computer. A specific application is determined to be monitored. The agent may be a user, using a learning application or another application they decide to track.

Machine learning researchers have discovered that certain types of feedback are highly beneficial for accelerating the rate at which reinforcement learning agents acquire complex motor skills (Ng, A. Y., Harada, D., & Russell, S. (1999). Policy invariance under reward transformations: Theory and application to reward shaping. In I. Bratko & S. Dzeroski (Eds.), Proceedings of the 16th Annual International Conference on Machine Learning (pp. 278-287). San Francisco: Morgan Kaufmann.) This line of work highlights that to be beneficial the additional feedback the agent is given has to obey to the principle of reward shaping. That is the feedback signal should be constructed as the difference between the (discounted) value of the new state and the value of the previous state. In our context, the value of different attentional states can be modelled in terms of the expected value of control theory. Instructional feedback is particularly effective when it aligns to learners’ achievement and prior knowledge and the complexity of the learning task. On this account, with increasing skill development, learners can benefit from fading instructional guidance to enable the transfer of the trained skill to real-life contexts. The solution presented herein uses a guidance fading out mode so that with increased progress in skill acquisition, learners get increasingly less explanations on the task. This avoids overloading the limited cognitive resources by unnecessary redundant information. The application to be monitored is preferably determined by the users themselves and is selected from a set of personal applications, the user wants to execute in his work. Thus, users can further benefit from personalized learning environments that contain individualized forms of address as social cues. Our approach builds on these findings and applies them to turn the application into a virtual educational companion. The related attribution of human characteristics to the technical system has been described by the process of anthropomorphism. To address core limitations of existing training systems, the solution provided in this application is to develop an attention training module that effectively trains people’s attention control skills in the real world. It might be applied to enhance students’ performance in online learning settings or making people more productive in the workplace. The core of our approach involves building an optimal metacognitive feedback mechanism that adapts to learners’ skill levels and individual needs. To achieve that, we build on computational models of cognitive control and task resumption and evaluate our training application in field studies.

On a neurophysiological level, the control session relates to brain activation patterns in regions such as the prefrontal cortex (RFC) and the anterior cingulate cortex (ACC). For details in relation to the LVOC model it is referred to: Lieder, F., Shenhav, A, Musslick, S., & Griffiths, T. L. (2018) Rational metareasoning and the plasticity of cognitive control. PLoS computational biology, 14(4), e 1006043. The LVOC model also relates to the allocation of cognitive control in the ACC. We build on the idea that people use reinforcement learning to predict the value of candidate control signals of different types and intensities based on stimulus features. Our model could predict performance patterns in a variety of experimental paradigms, including how people learn what to attend to in a visual search task. From the model, we derived an optimal feedback mechanism for promoting metacognitive reinforcement learning.

ACTrain is a Python-based software prototype, in the following also called method or application. In its basic functionality users can first decide for a task they want to work on subsequently, which avoids training with artificial tasks that lack relevance to real-life situations. Since people also specify related programs that will be used for this activity, the application can constantly track if the programs they use match the previously indicated list of programs and give feedback on that. The feedback emphasizes the previously set goals and reminds users of their own intention. After the initially proposed session duration expired, the performance in the current session is compared with the existing history, in particular with a weighted mean of the control level of past control sessions. Optionally, a Kalman filter may be used for these calculations. This enables users to keep track of their own progress in skill acquisition. To better adjust the application to users' individual needs, the provided support should align to the detected skill level. This can be achieved based on the guidance fading mode.

The core of this invention is to strengthen agent’s attention control skills in the real world and confirm this in field studies. To achieve that, we will enhance our current prototype of the computer-based attention training application with adaptive metacognitive feedback and turn it into a personalized virtual companion. On the technical side, the solution involves two feedback loops:

1. First, our computational model builds the theoretical foundation for developing an optimal feedback mechanism and inform subsequent improvements to the training program.

2. Second, as the software prototype is evaluated in user studies, the obtained human data can inform the computational modeling process and improve the model adaptively (self-learning mechanism) to provide further insights into how people's executive functions develop over time and the origin of individual differences in executive control. At first, an optimal feedback mechanism from the LVOC model is derived by applying the principle of reward shaping. The LVOC model predicts that reward shaping should effectively foster skill acquisition. This allows us to derive the value of control that informs the feedback signals. Concretely, our theory predicts that the reward the optimal feedback mechanism awards to the participant for their attention control at time tt should be a positive linear function of EVC(¾ c_t) - max EVC(¾, c), where EVC is the expected value of control as defined by

Shenhav et al. (2013) and ¾ and cj are the maximum likelihood estimates of the person’s attention control signal c_t and their internal state s_t respectively. The attention control signal c_t reflects which mental process (e.g., writing vs. reacting to a Facebook notification) the person boosted or inhibited and how strongly. And their internal state st reflects how strongly the corresponding neural pathways were activated prior to the control signal; this captures, for instance, how strong the urge to check Facebook was at that moment. The success of our optimal metacognitive feedback mechanisms at teaching people optimal planning strategies suggests that such feedback can also promote the development of strong attention control skills. The general idea to incorporate this principle into our application is to provide immediate rewards that accurately communicate the long-term value of keeping the attention focused on a previously set self-selected task goal. The application will track how people's attention control skill evolves over training sessions by using a Kalman filter and a generative model of how attention control skill and session related parameters (also called in short: session parameters), as session duration, affect people's ability to stay focused. Based on the inferred attention control skill level the system can provide intelligent feedback that accurately communicates the expected value of control (e.g., with a point-based reward system, color, facial expression, and sound). We will further implement a guidance fading procedure that reduces the frequency of feedback with increasing attention control skill level. It builds on evidence suggesting that students with higher levels of achievement can benefit more from delayed feedback. Second, the application will be adapted and personalized into a personalized virtual companion. The user has the opportunity to give the application a name and customize its appearance. Since it also addresses the user by name, they can build a personal relationship to the application. In line with existing research by fulfilling the need to connect with another social entity this dynamic can foster intrinsic motivation and lead to sustained use and goal-directed performance. The lab studies provide the opportunity to assess the usability of the application and cognitive demands that arise during the interaction with a set of different methods, including eye and mouse tracking, physiological parameters, and thinking aloud protocols. The method may be used by students and measure ACTrain’s effectiveness at improving students’ learning outcomes. Further, the method may be used by ADHD patients, since attention disorders are becoming increasingly more prevalent and can have dramatic effects in adult life.

To get further insights into how people's attention control skills develop overtime and the origin of individual differences in executive control, we will test the computational models of attention control skill acquisition against the real-life data. The method and the LVOC model, accordingly, is adaptively improved and extended to capture impaired executive functions related to ADHD, in particular behavioral inhibition, working memory, and self-regulation. Furthermore, a beneficial resumption strategy is used and supported by the application.

The method uses models for personalized interaction and adaptive user interfaces. Model- based user interface development (MBUID) is a predominant approach to adaptive interfaces for heterogeneous users and contexts of use. The basic idea behind MBUID for adaptive interfaces is that the model defines the interaction between user and system on a higher and abstract level. The concrete appearance and interaction mechanisms of the interface are not specified in the model - and thus, can be adapted in different variants of the final user interface. In many MBUID approaches, user interfaces are represented on several levels. Task and domain model, abstract Ul, concrete Ul and final Ul. This framework is very powerful and allows for extensive adaptations. However, adaptations of interface complexity and navigation structures to address attentional abilities would require significant modelling effort.

User, task and context profiling adaptivity requires reliable knowledge about the conditions to which the user interface shall adapt. This knowledge is usually collected in a user, task and context profile. As people with different attentional abilities can require very different adaptations to different user characteristics, there have been several efforts to develop user models for specific user groups, e.g. for motor-impaired users, for people with attentional deficits and for autistic users etc. A lot of work has been done in the field of context-aware systems. Ontology-based context models are very well established - an approach also followed by the MyUI project. Some researchers have proposed a taxonomy of context-aware software variability, but they do not specify the methods to detect context conditions. Especially for cognitive (attention) user conditions, temporal aspects play an important role. Cognitive states are changing dynamically whereas general user traits are relatively stable or change only very slowly over time. These variations should be captured by user, task and context models and indeed are considered in the feedback application which is suggested herein in order to manage this challenge effectively. The use of machine learning algorithms for user, task and context profiling has increased substantially in recent years, especially for recommender systems in the internet or to infer personal traits and preferences from interaction behaviors and activities in social networks. This idea can be transferred to user, task and context profiling where humans and computer can collaborate to create and populate profiles to adapt user interfaces Uls for optimized interaction and support for individual cognitive abilities of the user.

In particular, FIG. 1 is a schematic representation of a first computing unit 100 which is adapted for calculating and providing a formative feedback signal fbf on a Human-Machine-Interface HMI or a Human-Computer-Interface HCI, which may be provided on a monitor M or another technical device. Optionally, as shown in Fig. 1 with dotted lines, the formative feedback signal fbf is also provided to a second computing unit 200. The computing unit 100 comprises an input interface 101 for receiving a program signal pr_t, indicating whether the program that is currently in focus is congruent with the activity the user resolved to do in the current training session.

Further the input interface 101 is adapted for receiving input signals in. The input signals in may comprise:

Interaction signals is, relating to activity at input devices such as mouse or keyboard and confirms a focused state in the determined program according to the received signal pr_t.

EEG signals (electroencephalogram signals or categories thereof), fixation related potentials/FRP signals;

- Heart-rate related signals hs,

- fNIRS signals fNIRS, indicting brain activity in brain areas related to executive functions, e.g., the dorsal anterior cingulate cortex (dACC), the dorsolateral prefrontal cortex (DLPFC), or the ventrolateral prefrontal cortex (VLPFC).

Eye tracking signals af is used to measure whether the user is looking at specific goal- congruent or goal-incongruent areas of interest (a/_t) as an index of exogeneous components of attention.

Eye tracking is also used to measure the user’s pupil dilation pd_t as an index of endogenous components of attention (i.e., mental resource demands indicating cognitive involvement).

- The user’s galvanic skin response gs_t is used to estimate the user’s emotional involvement. - As shown in FIG. 1 with the dots in the last oval, also other input signals may be measured alternatively or in addition.

In other embodiment the list above is extendable by further signals (e.g. bio-signals) or may be limited to a specific set of input signals in. In a preferred embodiment, a first peripherphysiological signal (e.g. eye tracking) and a second neurophysiological signal (e.g. EEG etc.) are measured. The two multi-modal measurements are correlated.

In Fig. 1 the reference sign “COMP” relates to the evaluation whether a change in the attentional state st has been determined or not, as explained above (as-signal at time t deviates from earlier as-signal at time t-1).

FIG. 2 shows a second computing unit 200 with an input interface 201 and an output interface 202 for providing the summative feedback signal fbs on a Human-Machine-Interface HMI or a Human-Computer-Interface HCI, which may be provided on a monitor M or an another technical device. In particular, the estimated attentional state as_t (pr_t, is_t, ni_t, af_t, pd_t, gs_t, At_t) and situational variables par, such as the already passed time At_t, are then used to estimate the user’s attentional control skill sk_t(6, as_t,At_t ) - in FIG. 1 referenced with pred-sk - and an estimate of the control signal pred-c_t+i as formalized by the LVOC model according to Bayesian inference. At the end of a session, the improvement im_t is estimated from the sk_t and the prior skill estimate sk_t® (with sk₀ = 0_X) and used to update the prior skill estimate sk_t®. im_t informs the value V of the summative feedback signal fbs_t that is conveyed by a positive message if the system is at least 95% confident that the user’s attentional skill improved, shown in Fig. 2 with the reference sign IF (, where V is the value function of a Markov Decision Process model of the skill acquisition; see Xu, L, Wirzberger, M., & Lieder, F. (2019). How should we incentivize learning? An optimal feedback mechanism for educational games and online courses. In Proceedings of the 41st Annual Meeting of the Cognitive Science Society. Austin, TX: Cognitive Science Society). The summative feedback signal fbs_t may be provided on a monitor M.

FIG. 3 shows another embodiment, in which the two computing entities are integrated in one single computing unit, namely in a computer-implemented feedback module 1000 (in the following in short module). The module 1000 comprises the first and second computing entities 100, 200. It receives the input signals in and calculates and provides the formative feedback signal fbf on a monitor M and with other signals as, c to the second computing unit for calculating and providing the second feedback signal fbs which may again be provided on the monitor M. FIG. 4 is an interaction diagram, representing the message exchange between the respective entities, in particular between the memory MEM in which the LVOC model is stored, the computer-implemented feedback module 1000 and the agent or user. After a registration procedure and configuration phase, first, the application to be monitored is determined; this may be done by determining or receiving program signal pr (not shown in Fig. 4). Then, an input signal in may be measured (as mentioned above, like a neurophysiological signal etc.). In response to the input signal in, module 1000 calculates the attentional state as in steps S3, S4 and provides the formative feedback signal fbf to the user interface for each time. Thus, several formative feedback signals fbfu, fbfc, ...fbf_ti are sent to the monitor (due to transparency only two such signals are shown in Fig. 4). Further, based on the inferred attentional states as the memory is accessed in step S5 for using the LVOC model to calculate the summative feedback signal fbs in step S6 and provide the same to the module 1000, which may be forwarded in step S7 to the user/monitor M after the control session has ended, which is depicted in Fig. 4 with the dotted horizontal line. In another subsequent control session, the user has potentially improved his attentional skill level and provides “improved” input signals in’. All these signals are feed back to the LVOC model in order to adaptively improve the model itself, based on the provided signals in step S8.

FIG. 5 is a flow chart of the method for providing the feedback signals fbf, fbs. After start of the method, in step S1 the application to be monitored is determined (signal pr). In step S2 the input signals in are measured and provided to the computing unit 100, 1000 so that the attentional state as can be inferred in step S3. In step S4 the formative feedback signal fbf is calculated, based on the inferred attentional state as. The formative feedback signal fbf is only displayed on the monitor M or provided elsewhere, if a signal change has been detected to not disturb the user with unnecessary messages. After a pre-configurable period or according to pre-configured rules, the steps S1 to S4 are repeated so that several formative feedback signals are provided to the user within one control session in case a signal change has been observed. In step S5 the memory MEM is accessed to apply the LVOC model to the inferred actual attentional state as_t, the previous attentional state asu , the control signal, the feedback signals fbf, fbs and environmental or session related parameters par. In step S6 estimates are calculated, a prediction of the control signal pred-Ct_+i and a prediction for the skill level pred- sk. The skill level is only provided in case of a signal change. This should be represented in Fig. 2 with the query request IF. In step S7 an improvement signal is calculated so that the method may end or may be used to improve and adapt the mode with the predictions and the real measured values.

FIG. 6 shows an embodiment in which the first computing unit 100 and the second computing unit 200 are integrated into one computer-implemented feedback module 1000, which additionally comprises a determination unit D for receiving the signal pr indicating the application to be tracked. The module 1000 calculates and provides the two different feedback signals fbf, fbs by means of an output interface 202 that provides the signals fbs and fbf on a Human-Machine-Interface HMI or a Human-Computer-Interface HCI, which may be provided on a monitor M or an another technical device. In another embodiment, the first computing unit 100 and the second computing unit 200 may be implemented in one common unit or even may be identical. In the latter case, the two different functions/functionalities (calculating attentional states/formative feedback signal and calculation summative feedback signal) are executed on the same computing entity (e.g. processor etc.) The computer-implemented feedback module 1000 is adapted to execute the ACTrain method, which has been described above.

FIG. 7 is an overview figure for the architecture of the ACTrain method according to a preferred embodiment of the invention. As can be seen in Fig. 7, at each time point t in a control session, the system receives several multi-modal input signals. The signal pr_t indicates whether the program that is currently in focus is congruent with the activity the user resolved to do in the current training session.

ACTrain uses a Kalman filter KF to infer the user’s attentional state as_t from these signals, the estimate of the user’s attentional state in the previous time step as_t-l situational variables (e.g., how long the user has been in the current training session already, referred to as At_t), and an estimate of the user’s attention control skills Q. The estimation of as_t enables the system to provide formative feedback on the output interface 102 during a session by constantly monitoring if there has been a change from a state /, classified as focused, to a state d, classified as distracted, or vice versa. If a change was detected, a feedback signal fbf_/ = V(asi) - yCaSi i), where V is the value function of the rational metareasoning theory of cognitive control allocation, is emitted that is positive for (aS_j) - KCas_f-i) > 0 and negative for V (aS_[) - V{ asi-_j) < 0. Its output consists of visual components (color, facial expression, message), auditory components (sound), and a reward that that is proportional to the optimal feedback signal fbf_L. The estimation of the user’s or agent’s attentional control skill sk is calculated by the LVOC model according to Bayesian inference as has already been described with respect to Fig. 2 above. ACTrain is a general system for improving agent’s attention control skills in the real-world. Specific applications include the drug-free treatment of clinical and sub-clinical attention deficits in humans, including those that occur in ADHD.

Furthermore, ACTrain can be used to train and support workers who have to perform monitoring tasks that require high levels of focused attention and thus depend on solid and stable attention control skills. In safety critical operator domains, such as nuclear power plants or air traffic control, the emitted feedback signals can serve as a warning when a distracted operator state is detected. This facilitates re-focusing the attention to the task and prevents severe life-threatening accidents.

Any reference signs in the claims should not be construed as limiting the scope.

While the current invention has been described in relation to its preferred embodiments, it is to be understood that this description is for illustrative purposes only. For example, the agent may be a user or a non-human agent. Messages may also refer to digital messages that can be used in a broad set of business areas and applications such as audial, visual messages, information or education messages etc. For the person skilled in the art it is clear that the invention may be used for patients with attentional disorders as well as for healthy persons, who want to improve their attentional control skills. Also, the computing units need not to be deployed as physical servers. For example, it is also possible that the respective computing units described above can be hosted in a virtualized environment as well. Accordingly, it is intended that the invention be limited only by the scope of the claims appended hereto.

Wherever not already described explicitly, individual embodiments, or their individual aspects and features, described in relation to the drawings can be combined or exchanged with one another without limiting or widening the scope of the described invention, whenever such a combination or exchange is meaningful and in the sense of this invention. Advantageous which are described with respect to a particular embodiment of present invention or with respect to a particular figure are, wherever applicable, also advantages of other embodiments of the present invention.

Claims

Patent Claims

1. Computer-implemented method for providing a feedback signal during a user’s control session, comprising interactions with at least one application, based on measured signals, comprising:

- Determining (S1) the at least one technical application to be monitored;

- Measuring (S2) input signals (in), comprising a set of interaction signals from at least a group of sensors, embedded in a Human-Technology-Interface (HMI, HCI) which track a user’s interaction behavior during the control session;

- Using a computing unit (100) for continuously inferring (S3) attentional states (as) over time within the control session, from the measured input signals (in);

- Using the computing unit (100) for calculating (S4) a formative feedback signal (fbf), based on the inferred attentional states (as) to be provided on an output interface (102) in case a change of the attentional state (as) has been detected;

- Accessing (S5) a storage (MEM) with a stored model of plasticity of cognitive control (LVOC) and providing the inferred attentional states (as), a control signal (c) and the provided formative feedback signals (fbf) and session parameters (par) to the model (LVOC) for calculating (S6) an estimated control signal (pred-C_t+i) and an estimated attentional control skill level (pred-sk_t+i) by using a Bayesian inference algorithm based on a prior probability distribution of the attentional control skill level and the control signal;

- Using a computing unit (200) for calculating and providing (S7) an improvement signal as summative feedback signal (fbs) on an output interface (202) in case an improvement of the estimated attentional control skill level has been detected with a pre-configured confidence value.

2. Computer-implemented method according to claim 1, wherein the input signals (in) comprise:

- neurophysiological and/or peripherphysiological signals from a set of multi-modal neurophysiological and/or peripherphysiological sensors that can be attached directly to the human body or can be integrated into textile fibers, clothes, and/or elastic bands.

3. Computer-implemented method according to the directly preceding claim, wherein the neurophysiological and/or peripherphysiological signals are measured in reply to a sequence of formative and/or summative feedback signals (fbf, fbs) provided by the computing unit (100, 200).

4. Computer-implemented method according to any of the preceding claims 2 or 3, wherein the neurophysiological signals are measured by means of a functional near- infrared signal (fNIRS), representing human executive functions in pre-determined prefrontal cortical brain regions.

5. Computer-implemented method according to any of the preceding claims 2 to 4, wherein the neurophysiological signals are measured by means of an electroencephalogram (EEG) signal or categories thereof, comprising event-related potentials, oscillatory-based frequency measures, time-series and frequency-based connectivity measures, related to user’s attention and/or perceptual user feedback processes.

6. Computer-implemented method according to any of the preceding claims 2 to 5, wherein the peripherphysiological signals are measured by means of an eye-tracking (af) and/or pupillometry signal (pd), which is used by the processing unit (100) to estimate eye features such as fixation duration and pupil dilation activity related to user’s attention and/or perceptual user feedback processes.

7. Computer-implemented method according to the directly preceding claim, wherein the estimated eye-features, in particular a fixation onset, is used to estimate fixation-related categories of an EEG signal, such as fixation-related potentials related to higher order cognitive processes such as stimulus perception and allocation of attentional resources for controlling the application, or other types of input signals (in).

8. Computer-implemented method according to any of the preceding claims, wherein the peripherphysiological signal is measured by means of an electro-dermal signal (gs) and/or heart-rate signal (hs), related to user’s arousal and/or perceptual user feedback processes.

9. Computer-implemented method according to any of the preceding claims 2 to 8, wherein a set of features from the neurophysiological and/or peripherphysiological signals is selected, which adaptively optimizes inferring the attentional state of the user in a way that minimizes prediction error by selecting the sequence of optimal feedback signals using a Bayesian inference algorithm.

10. Computer-implemented method according to any of the preceding claims, wherein the set of interaction signals (is) for controlling the application comprises a window activation signal, indicating whether the window is currently used and to which application it belongs.

11. Computer-implemented method according to the directly preceding claim, wherein the window activation signal is generated by caching a window classifier from an operating system.

12. Computer-implemented method according to any of the preceding claims, wherein the control session is a learning skill task and wherein an adaption signal is provided for dynamically adapting a difficulty level of the task and/or a duration of the control session, based on the formative and/or summative feedback signal (fbf, fbs).

13. Computer-implemented method according to any of the preceding claims, wherein the formative feedback signals (fbf) are provided during execution of the control session in case a change of the attentional states (as) has been detected.

14. Computer-implemented method according to any of the preceding claims, wherein the summative feedback signal (fbs) is provided after execution of the control session and in particular after a pre-defined set of control sessions, in case a change of the summative feedback signal has been detected.

15. Computer-implemented method according to any of the preceding claims, wherein estimating an attentional control skill level comprises calculating a value function of a Markov Decision Process model of skill acquisition.

16. Computer-implemented method according to any of the preceding claims, wherein a guidance-fading-out-mode is used which adaptively controls a frequency and/or intensity for providing the formative feedback signal (fbf) and optionally also for providing the summative feedback signal (fbs), in response to the formative and/or summative feedback signal (fbf, fbs) and/or in response to pre-configured rules, stored in a rule database.

17. Computer-implemented method according to any of the preceding claims, wherein the processed signals and/or data, in particular the measured interaction signals (is), the measured set of multi-modal neurophysiological and peripherphysiological signals, the inferred attentional states (as), the formative feedback signal (fbf) and the summative feedback signal (fbs) are stored locally and/or wherein the method comprises generating an encrypted data package from the processed signals and/or data for preparing the encrypted data package to be securely sent to a server via a network connection.

18. Computer-implemented method according to any of the preceding claims, wherein the method comprises to support refocusing after distractions by teaching a resumption strategy based on the processed signals and/or data.

19. Computer-implemented method according to any of the preceding claims, wherein the formative feedback signal (fbf) and/or summative feedback signal (fbs) is/are provided on an output interface (102, 202), in particular on a user interface and may comprise a message in textual form, in particular a reward message and/or visual components and/or may be provided as sound signal and/or as control signal for controlling priority of the applications running in parallel with respect to the determined application, in particular a control signal for a z-order of a graphical user interface of the computer system.

20. Computer-implemented method according to any of the preceding claims, wherein a filter algorithm is used, in particular a Kalman filter, for the set of multi-modal neurophysiological and/or peripherphysiological sensors signals and/or for the interaction signals (is).

21. Using the computer-implemented method according to any of the preceding method claims or a computer-implemented feedback module for the improvement of attentional skills or for the treatment of attentional disorders selected from the group consisting of an attention deficit syndrome (ADHD) and other related diseases with a malfunction of a user's attention.

22. Computer-implemented feedback module (1000) for providing a feedback signal (fbf, fbs) during a user’s control session, comprising interactions with at least one application, based on measured signals, which is adapted to execute the method according to any of the preceding method claims, comprising:

- A determination unit (D) for determining the at least one application to be monitored;

- An input interface (101) which is adapted for measuring a set of input signals (in), comprising a set of interaction signals (is) from technical sensors, embedded in a Human-Technology-Interface which track the user’s interaction behavior during the control session;

- A processing unit (100) for continuously inferring attentional states (as), related to the control session, from the measured interaction signals (is);

- wherein the processing unit (100) is further adapted for providing a formative feedback signal (fbf), based on the inferred attentional state (as) on an output interface (102) in case a change of the attentional state has been detected; and

- wherein a processing unit (200) is adapted for accessing a storage (MEM) with a stored model of plasticity of cognitive control and for providing the inferred attentional states (as), a control signal (c) and the provided formative feedback signals (fbf) and session parameters to the model for calculating an estimated control signal and an estimated attentional control skill level by using a Bayesian inference based on a prior probability distribution of the attentional control skill level and the control signal; and

- wherein the processing unit (200) is adapted for providing an improvement signal as summative feedback signal (fbs) on an output interface (202) in case an improvement of the estimated attentional control skill level has been detected with a pre-configured confidence value;

- The output interface (102, 202) for providing the formative feedback signal (fbf) and the summative feedback signal (fbs).

23. Computer-implemented module according to the directly preceding claim, further comprising:

- A set of physiological interfaces which are adapted for receiving a set of multi-modal physiological signals from a set of multi-modal neurophysiological and/or peripherphysiological sensors, comprising brain signals, indicating brain activity patterns during the control session.

24. Computer program comprising program elements which induce a computer system to carry out the steps of the computer-implemented method for providing a feedback signal during a control session of a technical application according to one of the preceding method claims, when the program elements are loaded into a memory of the computer system.

25. Computer-readable medium on which program elements are stored that can be read and executed by a computer system, in order to perform steps of the method for providing an output signal as feedback during a control session of a technical application according to one of the preceding method claims, when the program elements are executed by the computer system.