Evaluating substances comprising odorants and / or flavourings

Abstract

The present invention provides a method of evaluating a substance comprising an odorant and / or a flavouring comprising obtaining neural response data of a subject's response to the substance; determining a first indication of the subject's arousal and / or valence response to the substance based on the neural response data; modifying the first indication to obtain a second indication of the subject's arousal and / or valence response by a method defined herein and evaluating the substance based on the second indication. The invention also provides apparatus and / or a non-transitory machine-readable medium which incorporate the method of the invention.

Description

EVALUATING SUBSTANCES COMPRISING ODORANTS AND / OR FLAVOURINGSFIELD OF INVENTION

[0001] The present invention relates to methods and apparatus for use in evaluating odorants and / or flavourings, preferably odorants.BACKGROUND

[0002] In humans, the evaluation of taste and smell is interlinked. For example, when tasting a food, a process called retro-nasal smelling usually occurs in which volatile compounds inside the mouth travel through the nasopharynx into the smell receptors, contributing to experience of flavour. Moreover, there is a significant overlap in how smell and flavour are processed in the human brain, resulting in similar behaviours and autonomic responses.

[0003] Odorants are a feature of many products such as household cleaning products (e.g. detergents, polishes and disinfectants), personal care compositions (e.g. soaps, shampoos, cosmetics and fragrances), food, packaging and the like. In many cases, odorants can add to the appeal of a product, for example by providing a pleasant scent, evoking a particular association and / or masking an underlying less pleasant odour.

[0004] Flavourings include products and ingredients which are included in foods or other substances which are intended to be tasted. The term ‘flavouring’ can include odorants due to the close link between taste and smell. In some examples, flavourings are intended to give food a particular or stronger taste and / or smell. Flavours may for example have their origin in plants, animals or microbiological entities, and / or may be synthesised.

[0005] In some examples, odorants and / or flavourings are evaluated in order to determine whether they meet predetermined criteria. In some examples, the predetermined criteria may be an “emotional target”, which describes a human subject’s likely or average response to the odorant / flavouring. This may be characterised as a combination of arousal and valence values. In this context, “arousal” comprises thedegree of activation or intensity that accompanies an emotional state whereas “valence” comprises different pleasantness states and differentiates positive from negative emotions.

[0006] It will be appreciated that there is a wide range of such “emotional targets” and a particular target may depend on an intended use case. In some examples, it may be intended to identify an odour / flavour which makes a product pleasant to use or consume whereas in other examples an unpleasant odour / flavour may be deliberately sought to cause users to avoid consuming the product which may be harmful to them. Odours and flavours may be associated with perception of a value of the product. In some examples, a manufacturer may want to replace at least one ingredient comprising or within an odour / flavour creating component (i.e. an odorant or a flavouring) of their product. For example, an existing ingredient may not be approved or appropriate for use in certain geographical jurisdictions or may have become difficult or expensive to obtain. However, the manufacturer may wish to maintain a consistent response to their product and in such cases, it may be intended that the new ingredient, alone or in combination with other ingredients, evokes a response which matches the response for the product made using the ingredient which is to be replaced.

[0007] A subject’s response to an odorant / flavouring may be evaluated using a survey. For example, an emotional state questionnaire using a Self-Assessment Manikin (SAM) scale to measure valence and arousal may be used. However, individual responses can be highly subjective and can vary significantly even for a single subject over the course of a day.

[0008] Some methods have been developed to reduce the subjectivity of reporting a response by relying on measurable biometric or physiological responses to an odorant. For example, EP2799864 in the name of Procter & Gamble (incorporated herein by reference) discloses a method of determining a biometric fingerprint for a fragrance composition according to biometric response data measuring various Autonomic Nervous System (ANS) parameters. The parameters may include electrodermal indices, such as skin conductance, and thermovascular indices such as skin temperature (ST), respiratory rate (RR), heart rate (HR), and blood pressure (BP diastolic & systolic).

[0009] In another example, US9295806 in the name of iMotions AS (incorporated herein by reference) discloses a method of determining emotional response to an olfactory stimulation by collecting and analysing eye data including pupil data, blink dataand gaze data. The authors suggest that this may include partially pre-cognitive responses as well as a cognitive emotional response.SUMMARY OF INVENTION

[0010] According to a first aspect, a method of evaluating a substance comprising an odorant or flavouring comprises obtaining neural response data of a subject’s response to the substance and determining a first indication of the subject’s arousal and / or valence response to the substance based on the neural response data, for example based on a predetermined relationship and / or a trained model. The method may further comprise obtaining at least one of (i) an arousal response profile for the subject based on autonomic nervous system biometric data obtained in response to the substance to be evaluated, and (ii) a valence response profile for the subject based on facial analysis of images acquired while the subject was exposed to the substance. The first indication of the subject’s arousal and / or valence response to the substance is modified based on at least one of the arousal response profile and the valence response profile to obtain a second indication of the subject’s arousal and / or valence response, and the substance is evaluated based on the second indication. In particular, if the first indication is an indication of the subject’s arousal response (and not the subject’s valence response), or an indication of the subject’s arousal and valence response, modifying the first indication may comprise modifying the first indication based on the arousal response profile. If the first indication is an indication of the subject’s valence response (and not the subject’s arousal response), or an indication of the subject’s arousal and valence response, modifying the first indication may comprise modifying the first indication based on the valence response profile.

[0011] Neural response data (for example comprising, or derived from, electroencephalography (EEG) data) may be indicative of a subject’s valence and / or arousal response to a substance comprising an odorant and / or a flavouring. However, interpretation of such data is complex. By modifying this data with an arousal response profile and / or a valence response profile, a more accurate determination of a subject’s response to the substance may be made.

[0012] According to a second aspect, processing circuitry comprises a neural response analysis module, a response modification module and an evaluation module.The neural response analysis module is configured to obtain neural response data of a subject’s response to a substance comprising an odorant or flavouring and determine a first indication of the subject’s arousal and / or valence response to a substance based on the neural response data, for example based on a predetermined relationship or a trained model. The response modification module is configured to obtain an arousal response profile for the subject based on autonomic nervous system biometric data obtained in response to the substance to be evaluated, and / or to obtain a valence response profile for the subject based on facial analysis of images acquired while the subject was exposed to the substance and to modify the first indication of the subject’s arousal and valence response to the substance based on at least one of the arousal response profile and the valence response profile to obtain a second indication of the subject’s arousal and / or valence response. The evaluation module is configured to evaluate the substance based on the second indication. In particular, if the first indication is an indication of the subject’s arousal response (and not the subject’s valence response), or an indication of the subject’s arousal and valence response, modifying the first indication may comprise modifying the first indication based on the arousal response profile. If the first indication is an indication of the subject’s valence response (and not the subject’s arousal response), or an indication of the subject’s arousal and valence response, modifying the first indication may comprise modifying the first indication based on the valence response profile.

[0013] According to the third aspect, a machine-readable medium comprises instructions which when executed by a processing circuitry cause the processing circuitry to obtain neural response data of a subject’s response to a substance comprising an odorant or flavouring and determine a first indication of the subject’s arousal and / or valence response to the substance based on the neural response data, for example based on a predetermined relationship, e.g. a trained model. The instructions further comprise instructions to cause the processing circuitry to obtain at least one of (i) an arousal response profile for the subject based on autonomic nervous system biometric data obtained in response to the substance to be evaluated and (ii) a valence response profile for the subject based on facial analysis of images acquired while the subject was exposed to the substance; and to modify the first indication of the subject’s arousal and valence response to the substance based on at least one of the arousal response profile and the valence response profile to obtain a second indication of the subject’s arousaland / or valence response. The instructions further cause the processing circuitry to evaluate the substance based on the second indication. In particular, if the first indication is an indication of the subject’s arousal response (and not the subject’s valence response), or an indication of the subject’s arousal and valence response, the instructions to modify the first indication may comprise instructions to modify the first indication based on the arousal response profile. If the first indication is an indication of the subject’s valence response (and not the subject’s arousal response), or an indication of the subject’s arousal and valence response, instructions to modify the first indication may comprise instructions to modify the first indication based on the valence response profile.DESCRIPTION OF FIGURES

[0014] Examples will now be described in conjunction with the following drawings in which:

[0015] Figure 1 shows a schematic representation of processing circuitry;

[0016] Figure 2 shows a flowchart of an example method for evaluating a substance comprising an odorant or flavouring;

[0017] Figure 3 shows an example method of determining a response profile to a substance comprising an odorant or flavouring;

[0018] Figure 4A shows an example of a first indication of a subject’s arousal and valence;

[0019] Figures 4B and 4C show example response profiles;

[0020] Figure 4D shows an example of a modified arousal and valence response;

[0021] Figure 5 shows an example of processing circuitry in conjunction with a machine readable medium; and

[0022] Figure 6 shows a system for determining a subject’s response to an odorant.DETAILED DESCRIPTION

[0023] Figure 1 shows an example apparatus comprising processing circuitry 100 which may be used in evaluating a substance comprising an odorant or flavouring. Preferably the substance comprises or is an odorant. The substance may comprise orbe a flavouring. The term ‘odorant’ applies to any chemical or chemical composition which can evoke an olfactory response in a human subject. The term ‘flavouring’ applies to any chemical or chemical composition which can evoke a taste response in a human subject.

[0024] Odorants or flavourings may comprise at least one fragrance or flavour material. Preferably the odorant comprises a fragrance material. The nature and type of the fragrance or flavour materials do not warrant a detailed description here, which in any case would not be exhaustive, the skilled person being able to select them on the basis of its general knowledge and according to the intended use or application and the desired organoleptic effect. Typically, a fragrance material is a mixture of fragrance compounds. The fragrance compounds may belong to chemical classes as varied as alcohols, aldehydes, ketones, esters, ethers, acetates, nitriles, terpenoids, nitrogenous or sulphurous heterocyclic compounds and essential oils, and said materials can be of natural or synthetic origin. Many of these fragrance compounds are listed in reference texts such as the book by S. Arctander, Perfume and Flavor Chemicals, 1969, Montclair, New Jersey, USA, or its more recent versions, the relevant parts of which are incorporated herein by reference.

[0025] Fragrance materials may comprise at least one fragrance compound selected from: i) hydrocarbons; ii) aliphatic alcohols; iii) aliphatic ketones and oximes thereof; iv) aliphatic carboxylic acids and esters thereof; v) acyclic terpene alcohols; vi) acyclic terpene aldehydes and ketones; vii) cyclic terpene alcohols; viii) cyclic terpene aldehydes and ketones; ix) cyclic alcohols; x) cycloaliphatic alcohols; xi) cyclic and cycloaliphatic ethers; xii) (ethoxymethoxy)cyclododecane; xiii) cyclic ketones; xv) esters of cyclic alcohols;xvi) esters of cycloaliphatic carboxylic acids; xvii) aromatic and aliphatic alcohols; xviii) esters of aliphatic alcohols and aliphatic carboxylic acids; xix) aromatic and aliphatic aldehydes; xx) aromatic and aliphatic ketones; xxi) aromatic and aliphatic carboxylic acids and esters thereof; xxii) nitrogen-containing aromatic compounds; xxiii) phenols, phenyl ethers and phenyl esters; xxiv) heterocyclic compounds; xxv) lactones; and xxvi) essential oils.

[0026] The fragrance material may comprise at least one solvent. The solvent may be a hydrophobic material that is miscible with the fragrance compounds. The solvent may provide at least one of the following benefits: i) increase the compatibility of compounds in the fragrance material, ii) increase the overall hydrophobicity of the fragrance material, iii) influence the vapor pressure of the fragrance material, and iv) provide rheological structure to the fragrance material. Suitable solvents are those having reasonable affinity for the fragrance compounds. The affinity may be determined by using a group contribution method to predict a partition co-efficient which can be expressed by a ClogP value.

[0027] The processing circuitry 100 comprises a neural response analysis module 102, a response modification module 104 and an evaluation module 112. In use of the processing circuitry 100, neural response data 106 is obtained by the neural response analysis module 102. The neural response analysis module 102 determines a first indication of a subject’s arousal and / or valence response to an odour, for example using a predetermined relationship, which may comprise a trained model. The response modification module 104 is configured to modify the first indication based on at least one of autonomic nervous system biometric data 108 and facial analysis data 110, as will be described in greater detail below.

[0028] In an example, the neural response data 106 may comprise, or be derived from, electroencephalography (EEG) data. EEG is a technique used to measure brain electrical activity. For example, electrodes may be attached or placed upon a user’s head (for example, being arranged within a headset or skullcap) and used to recordbioelectrical signals generated by the brain in response to stimuli. In some examples, the raw electrical signals may be processed to extract features or parameters thereof. For example, EEG data may be processed to extract any or any combination of Hjorth parameters, spectral entropy, measures of energy, frontal alpha asymmetry, spectral slope, Differential Alpha Asymmetry, a ratio between channels or the like, as is described in greater detail below. In an EEG, a channel represents the difference between electrical levels measured by two electrodes. In some examples, a ratio between frontal channels (i.e. , channels monitoring the frontal lobe of the subject) may be determined. Frontal channels may be for example as defined in the international “10 - 20” system, or the 10% division thereof (also referred to as the 10-10 system) for placing electrodes for EEG recordings. For example, using the 10-10 system, a ratio may be defined as a ratio between a sum of the beta signal and an alpha signal from each of a plurality of frontal channels. In particular, a ratio may be defined as a ratio between (i) a sum of the power of the beta waves and (ii) a sum of the power of the alpha waves from each of a plurality of frontal channels. In such a context, ‘power’ is used to assess which frequencies are dominant for a specific signal in time. Higher power for a range of frequencies (e.g. beta waves) would mean that those are dominant over others with a different range.

[0029] In some examples, the neural response analysis module 102 may obtain the neural response data by extracting features from EEG data. In other examples, features may be extracted prior to the data being provided to the neural response analysis module 102. The neural response analysis module 102 in this example is configured to process the neural response data 106 according to a relationship. In some examples, the relationship may be embodied as at least one learnt model, for example model(s) learnt using machine learning techniques.

[0030] In an example, at least one machine learning model can be trained using neural response data (e.g. EEG data, or data derived from EEGs) which is associated with survey data indicating the response of subjects to a stimulus. In some examples, the stimulus is an olfactory or taste stimulus, provided by an odorant or flavouring and the machine learning model is trained using neural response data comprising EEG data, or data derived from an EEG, which is paired with survey response data (e.g. SAM survey responses). Examples of developing models are set out in greater detail below. In some examples, models may be trained to output an indication of valence, arousal or a combination thereof.

[0031] An olfactory stimulus may be provided by causing a subject to smell an ingredient or composition which comprises or contributes to an odorant such as a fragrance material, masking odorant or the like. In some examples, molecules of the ingredient(s) may be transported through an air flux generated by an air pump and controlled by an olfactometer. An olfactometer is an apparatus which may include a control unit to manage air flow given to an olfactory outlet. The air pump may be used to generate an air flow which is then mixed with air carrying molecules of the odorant.

[0032] A taste stimulus may be provided in a number of ways. In some examples, a flavour is provided in a liquid form and / or under controlled lighting so as to minimise the impact of texture and / or visual stimuli in the subject’s response. For example, flavourings may be tested by being drunk from a coloured container (e.g. an amber glass bottle) under a coloured light (e.g. a red light). Moreover, as will be described in greater detail below, in some examples, it may be intended to limit facial movements associated with chewing. While such conditions may be particularly suited to isolating the subject’s response to the flavouring, the test environment may be less rigorous in other examples. For example, foodstuffs may be tasted in the form that they are intended to be consumed.

[0033] In examples, the target output of the trained model is a predicted survey response for the subject. In some examples, this may comprise a single classification of their arousal and / or valence. However, in other examples herein, the output may comprise a matrix indicative of arousal and / or valence for a particular user in response to a particular odorant / flavouring. Such a matrix may comprise probabilities, based on input EEG data of a user exposed to a particular odorant, that the valence and the arousal of the user corresponds to each possible combination of valence and arousal values, or to individual valence and arousal values. In some examples (in particular corresponding to SAM questionnaire responses), there are 5 such values (-2, -1 , 0, 1 , and 2) for each of valence and arousal. Thus the output may provide a 5 by 5 matrix of probability values. In other examples, the output may comprise a valence vector of probability values which are associated with a valence response (e.g. a 1 by 5 vector of probability values), and / or an arousal vector of probability values which are associated with an arousal response (e.g. a 1 by 5 vector of probability values). In other words, the model may be trained to act as a probability, or probabilistic, classifier to infer or estimate valance and / or arousal responses of a subject from EEG data.

[0034] If the substance was to be evaluated based on such matrix / matrices or vector(s), or any assessment based purely on neural response data, then the valence and / or arousal value associated with the highest probability value could be selected, and used to classify the substance as having a corresponding valence and / or arousal value. However, in examples herein, the output is further processed using autonomic nervous system biometric data 108 which is indicative of arousal, and / or facial analysis data 110 relating to the face of the subject when exposed to the substance, which is indicative of valence.

[0035] In particular, the biometric data and / or the facial analysis data may each or both be used to modify the initial response determined using neural response data. This data may separately reflect arousal and valence, which in turn may be used to clarify the result of the neural response data. Moreover, the inclusion of neural response data may help to clarify individual responses from subjects (e.g. clarify a valence response of an individual with relatively non-expressive facial expressions).

[0036] In a particular embodiment, biometric data may be collected data which is indicative of arousal whereas facial analysis is indicative of valence. The biometric data and / or the facial analysis data may be collected in an overlapping timeframe with the neural response data 106. In other examples, they may be collected in a same session from a same user, for example within an hour, or within half an hour, or within 10 minutes of each other. In some examples, the neural response data may be acquired for a portion of the time in which the biometric data and the facial analysis data is collected. This is because, as outlined below, the biometric data and facial analysis data may be acquired while a subject responds to multiple substances including a target substance (i.e. a substance which is to be evaluated by the apparatus and method set out herein), whereas the neural response data 106 may be acquired in relation to just the target substance. In such examples, the biometric data 108 and / or the facial analysis data 110 for the target substance may be acquired at the same time as, or in an overlapping timeframe with, the neural response data 106 for the target substance. In some preferred examples, when the substance is an odorant, steps may be taken to exclude subjects who have anosmia, hyperosmia, (or related conditions which impact the sense of smell), and / or people who use their sense of smell for a living. Similar steps may be taken when evaluating a flavouring. In such cases, any or all of people who use their sense of taste for a living, people with ageusia and / or so-called ‘supertasters’ may be excluded. In someexamples, the subjects may be taken from a population made up of a target market for a product which may be associated with the odorant / flavouring.

[0037] The ANS comprises the regulation of involuntary physiological processes such as heart rate, skin conductance, respiration, and digestion. Physiological processes may be regarded as ways of measuring arousal.

[0038] Examples of biometric data 108 which are indicative of arousal may comprise any or any combination of parameters such as skin conductance (which may be variously referred to as galvanic skin response (GSR), electrodermal response (EDR), psychogalvanic reflex (PGR), skin conductance response (SCR) or skin conductance level (SCL)), skin temperature, respiratory rate, heart rate, blood pressure, pulse oximetry , pupillometry, electrogastography or the like. These are all measures of an autonomic nervous system response. Other measures of autonomic nervous system (ANS) response may be used in other examples.

[0039] Skin conductance is a method of measuring the electrical conductance of the skin, which varies with its moisture level, and may be measured in Siemens per meter or the like. The sweat glands are controlled by the sympathetic nervous system, and thus skin conductance may be used as an indication of psychological or physiological arousal.

[0040] Skin temperature is the response of the ANS to changes in psychological state and can be detected through measurement of signals reflecting dermal activity and / or using temperature sensors. Skin temperature may for example be extracted from a thermal image or measured directly, e.g. using a skin thermistor.

[0041] Respiratory rate (RR, Vf or Rf) is also known as respiration rate, pulmonary ventilation rate, ventilation rate, or breathing frequency and is the number of breaths taken within a set amount of time, for example 60 seconds.

[0042] Heart rate (HR) is the number of heart beats per unit time, usually expressed as “beats per minute” (bpm).

[0043] Blood pressure (diastolic and / or systolic) refers to the arterial pressure of the systemic circulation. During each heartbeat, blood pressure varies between a maximum (systolic) and a minimum (diastolic) pressure. The blood pressure in the circulation is principally due to the pumping action of the heart. The measurement of blood pressure usually refers to the systemic arterial pressure measured at the upper arm and is a measure of the pressure in the brachial artery, a major artery in the upper arm. Bloodpressure is usually expressed in terms of the systolic pressure over diastolic pressure and is measured in millimetres of mercury (mmHg).

[0044] Electrogastography is a non-invasive technique used to record gastric myoelectrical activity, measured by placing electrodes on the abdominal skin close to the stomach and which can identify an impact on stomach rhythms as may arise from the brain-gut connection.

[0045] Pupillometry is the measurement of pupil size and reactivity, which has been shown to change in response to arousal or emotion.

[0046] Pulse oximetry may be measured using a pulse oximeter and uses a light emitting diode and light detector to measure blood oxygen saturation, i.e. a percentage of oxygen attached to haemoglobin in the blood. Oxygenated blood takes in more infrared light, while deoxygenated blood takes in more red light. This difference is perceived by the absorption rate in the detector and can be converted into a percentage of oxygen in the blood.

[0047] In particular examples herein, a combination of electrocardiogram (ECG) data and electrodermal response (EDR) is used to determine an arousal level for a subject. Such measures are non-invasive and relatively easy to acquire. An ECG measures heart electrical activity. For example, an ECG may utilise three-leads with a positive, negative, and a ground electrode which are attached to a subject’s body. EDR may for example be measured using two electrodes, which may be placed on or near a subject’s hand or wrist as there are generally high sensitivity sweat glands in that area. The variation of sweat gland activity may be described as a variation of the skin conductance. EDR may in particular be a good indication of the actual response to the stimulus and relatively free of contamination from other bodily processes (e.g. digestion and the like) which may impact ECG data.

[0048] However, in other examples just one of these measures or any or any combination of other measures of ANS response may be used. In particular, other measures which are relatively easy to acquire include respiratory rate and / or blood pressure.

[0049] Facial analysis data may include facial expression analysis while experiencing a stimulus. In examples herein, facial expressions may be captured by a camera, for example as a video, while the subject is experiencing an olfactory stimulus or consuming a substance. In addition, in some examples, features may be extractedfrom one or more images captured by the camera. In some examples, image analysis is applied to camera or video images. Examples of extracted features which may be utilised alone or in any combination in methods herein include: brow furrows; eye lid tightening; eye movement (e.g. variation in direction of gaze); mouth movement (for example, smile recognition); blink frequency; nose wrinkle; cheek raise; chin raise and the like. Such facial states or movements may be extracted from a sequence of images by tracking the movement of features identified in the image. In some examples, a number of times that a subject’s face assumed a particular state or performed a movement may be counted. In some examples, measures which result in a high degree of variability for a particular subject and / or group of subjects may be selected. In other words, the response to different substances by at least some subjects may be more marked for certain indicators than for others, and such indicators may be selected to provide a subject’s valence profile. In some examples, facial analysis may instead provide an indication of emotion such as Anger, Contempt, Disgust, Fear, Joy, Sadness, Surprise, Engagement, Valence, Sentimentality, Confusion, Neutral, Attention, based on predetermined patterns. For example, this may utilise artificial intelligence techniques and models to identify a likely emotion.

[0050] In examples herein, each of the biometric data and the facial analysis data may be used to develop, respectively, a subject’s arousal response profile and valence response profile to the substance under test. In some examples, the arousal response profile and / or the valence response profile may be evaluated relative to a range of responses of the subject to each of a set of substances. For example, the set of substances may provoke a range of responses, and an average (e.g. median) and measure of variability (e.g. standard deviation) may be determined. A response scale may be defined based on the average and the measure of variability. The position of the response to the substance to be evaluated on the scale may be determined and used in turn to determine the response profile. In some examples, the response profile may be defined as a plurality of probability values associated with response levels (e.g. SAM questionnaire response levels), for example expressed as a vector. In some examples, response profiles may be predetermined and selected based on the response to the substance to be evaluated. An example of determining a response profile is described in greater detail below with reference to Figures 3 and 4A to 4D.

[0051] Each response profile may then be used to modify the output of the neural response analysis module (i.e. the first indication of the subject’s arousal and / or valence response to the substance). For example, where the first indication of the subject’s arousal and / or valence response is a probability matrix, this probability matrix may be modified using the arousal response profile and / or the valence response profile. For example, a value for a cell in the matrix may be replaced with an average of the original value, the value from a valence response vector associated with that cell and the value from an arousal response vector associated with that cell. This may adjust the probabilities in the matrix. In some examples, this may mean that, following modification, the most likely valence and / or arousal state is different than in the first indication of the subject’s arousal and / or valence response. If the first indication was a predicted estimated arousal and / or valence state, this may be modified using the corresponding response profile(s) (i.e. a first indication of an arousal response may be modified using an arousal response profile and a first indication of a valence response may be modified using a valence response profile), for example being increased or decreased depending on whether the estimation made in the response profiles is for a higher or lower arousal / valence than that estimated in the first indication.

[0052] The second indication of the subject’s arousal and / or valence response is then evaluated by an evaluation module 112 which produces an evaluation of the substance.

[0053] In some examples, an evaluation may be produced for each of a plurality of subjects’ responses to the substance in a similar manner and the substance may be evaluated based on an aggregate response (for example based on an average response). In some examples, based on the evaluation, the substance may be selected or recommended for inclusion in a product.

[0054] For example, the product may comprise a foodstuff, a personal care formulation or a home care formulation.

[0055] The term ‘personal care formulation’ as used herein means a consumer product intended to be applied to the human body or any part thereof for cleansing, beautifying, or improving appearance. Personal care formulations include but are not limited to cosmetics; deodorants; bar soaps; liquid soaps; facial and body washes; facial and body cleansers; hair shampoos; hair conditioners; toothpastes; shaving creams orgels; and foot care products. A personal care formulation does not include any product for which a prescription is required.

[0056] The term ‘home care formulation’ when used herein means a consumer product for use by household and / or institutional consumers for cleaning, caring, or conditioning of the home. Home care formulations include but are not limited to detergents including laundry detergents and dishwashing detergents; conditioners including fabric conditioners; cleaning formulations including hard surface cleaners; polishes and floor finishes.

[0057] Figure 2 is an example of a method of evaluating a substance, which may be carried out by processing circuitry, for example by the processing circuitry 100 of Figure 1.

[0058] The method comprises, in block 202, obtaining neural response data of a subject’s response to the substance to be evaluated. For example, as set out above, this may comprise acquiring EEG data (for example, directly from a user by carrying out an EEG to measure brain electrical activity, over a network or from a memory), or features extracted therefrom (for example, over a network or from a memory). In examples where EEG data is acquired, the method may comprise extracting features therefrom.

[0059] Block 204 comprises determining a first indication of the subject’s arousal and / or valence response to the substance based on the neural response data, for example based on a predetermined relationship. For example, this may comprise an estimated arousal and / or valence response. In some examples, the first indication may comprise a probability matrix or vector wherein a plurality of arousal and / or a plurality of valence levels are associated with a probability. In some examples, determining the first indication utilises a trained model or models, wherein the model(s) have been trained using machine learning or artificial intelligence principles. In some examples, two models may be trained, one in relation to valence and one in relation to arousal. In some such examples, the output of such models may be combined, e.g. averaged, to provide a combined indication of valence and arousal. In the example of Figure 2, the first indication is a first indication of the subject’s arousal and valence response, but just one of these may be evaluated in other examples.

[0060] Block 206 comprises obtaining an arousal response profile for the subject based on autonomic nervous system biometric data obtained in response to the substance to be evaluated. For example, this may comprise obtaining (which in someexamples comprises measuring, directly from a subject) data indicative of any or any combination of an electrocardiography, skin conductance, skin temperature, respiratory rate, heart rate, blood pressure, pulse oximetry, pupillometry, electrogastography or any other data indicative of an ANS response of the subject. In some examples, a subject’s arousal response is obtained to a set of substances (e.g., the subject may be monitored while being exposed to the set of substances) and the arousal response profile is determined relative to the range of arousal responses provided by the subject.

[0061] Block 208 comprises determining a valence response profile for the subject based on facial analysis based on images of the subject’s face acquired while the subject was exposed to the substance. For example, this may comprise obtaining data indicative of at least one aspect of a user’s facial expression while exposed to the substance. In some examples, a subject’s valence response is obtained to a set of substances (e.g. the subject may be filmed or photographed while being exposed to the set of substances, and in some examples, the method may comprise filming and / or photographing the subject) and the valence response profile is determined relative to the range of valence responses provided by the subject.

[0062] While in this example, both blocks 206 and 208 are shown, in some examples only one of the blocks may be carried out. Moreover, the data may be acquired directly from a user (i.e. measured or captured as an image), or from a memory, over a network or the like.

[0063] Block 210 comprises modifying the first indication of the subject’s arousal and / or valence response to the substance based on at least one of the arousal response profile and the valence response profile to obtain a second indication of the subject’s arousal and / or valence response. For example, where the first indication is a probability matrix, the probabilities therein may be modified based on the arousal response profile and / or the valence response profile. In some examples, the arousal response profile and / or the valence response profile may comprise probability vectors which may be combined with the probability matrix of the first indication.

[0064] For example, if the first indication is an indication of the subject’s arousal response (without an indication of the subject’s valence response), or an indication of the subject’s arousal and valence response, block 210 may comprise modifying the first indication of the subject’s arousal and / or valence response to the substance based on the arousal response profile. If the first indication is an indication of the subject’s valenceresponse (without an indication of the subject’s arousal response), or an indication of the subject’s arousal and valence response, block 210 may comprise modifying the first indication of the subject’s arousal and / or valence response to the substance based on the valence response profile. If the first indication is indicative of both the arousal and valence response of the subject, then block 210 may comprise modifying the first indication based on the arousal response profile and / or the valence response profile.

[0065] Block 212 comprises evaluating the substance based on the second indication. For example, this may comprise selecting a most probable value from a modified probability matrix as indicating the valence and / or arousal response to the substance. In some examples, the evaluation may comprise comparing the most probable value from a modified probability matrix with a target classification value and the substance may be evaluated based on the difference between the most probable value and the target classification value.

[0066] Figure 3 shows an example of a method for determining an arousal / valence response profile.

[0067] Block 302 comprises obtaining a plurality of responses of a particular subject to each of a set of substances.

[0068] For example, when determining an arousal response profile, this may comprise acquiring data indicative of, or derived from, electrocardiography data of a subject as the subject was subjected to different substances. This may comprise determining an average heart rate and / or an average R-R interval wherein an R-R interval is a time that elapses between two successive R waves of a “QRS” signal of an electrocardiogram. As a further alternative, a subject’s dermal response may be determined. For example, a subject’s skin conductivity may be determined while they are in a resting state and a number and / or amplitude of peaks in skin conductance relative to the resting state may be determined in response to each of the plurality of substances. In still further examples, a number or interval of breaths may be determined. Other parameters indicative of an autonomic nervous system response may be acquired in other examples.

[0069] When determining a valence response profile, block 302 may comprise acquiring facial analysis data which may count, and / or may evaluate a magnitude, of facial movements or states. For example, a count of facial expression states such as a number of brow furrows or lid tightenings may be counted for at least part of the periodof time for which a subject is exposed to a substance. In other examples, facial analysis algorithms which estimate emotions of the subject may be utilised.

[0070] Where the substance is an odorant, in examples, the time period for which a subject is subjected to each odorant may be consistent for example being in the region of between around 10 seconds to 1 minute, or around 30 seconds.

[0071] In some examples, exposing a subject to an odorant comprises supplying the odour using an olfactometer as described above. In other examples, a subject may be asked to smell a scent pad treated with an odorant, or the odour may be delivered in some other way.

[0072] Where the substance is a flavouring, in some examples, the flavour may be consumed or tasted in a consistent way, for example being consumed in liquid form and / or from particular containers, in some examples under consistent lighting conditions. In other examples, the flavourings may be delivered on a wafer or another similar consistent substrate. As in examples herein facial analysis may be carried out, the delivery system may be selected so as to minimise facial movements such as chewing, as these may otherwise add to the complexity of interpreting the subject’s response. In addition, facial movements can impart artifacts on EEG data as electrodes and / or a skullcap or the like may be caused to shift due to such movements. Liquids, or dissolvable substrates such as wafers (which may dissolve in the mouth) or the like may provide suitable means for delivering flavourings without requiring the subject to chew.

[0073] While the method may be carried out locally to the subject, this need not be the case in all examples. In some examples, the data gathered from the subject may be transferred over a network or the like and / or stored in a memory prior to the method being performed. In some examples, the method may comprise obtaining the data, by acquiring images and / or biometric readings of the subject.

[0074] Once a set of readings for a particular parameter has been acquired, block 304 comprises determining an average and a variability of the set of readings. In one example, this may comprise determining a median value as well as a standard deviation. The response to the substance may then be evaluated based on a scale determined using the average and the standard deviation.

[0075] In this example, in block 306, one of a set of predetermined response profile vectors may be selected based on a numerical difference between the value for a parameter measured in response to the substance to be evaluated and valuesdetermined based on the average and the measure of variability. For example, if the detected value for a parameter is closest to a value of the median less two standard deviations, then a first vector is selected. If the detected value for a parameter is closest to a value of the median less one standard deviation, then a second vector is selected. If the detected value is closest to the median, then a third vector is selected. If the detected value for the parameter is closest to the median plus one standard deviation, then a fourth vector is selected and if the detected value for the parameter is closest to the median plus two standard deviations, then a fifth vector is selected.

[0076] For example, assuming that the parameter being measured is associated with an increased value for high arousal and / or an indication of likeability and the vectors represent probabilities associated with the 5 states of a SAM scale (-2, -1 , 0, 1 , 2), the vectors may be defined as follows:1stvector [0.4 / / 0.3 / / 0.2 / / 0.1 / / 0.0] (highly relaxing / highly distasteful)2ndvector [0.3 / / 0.3 / / 0.2 / / 0.1 / / 0.1] (Relaxing / distasteful)3rdvector [0.2 / / 0.2 / / 0.2 / / 0.2 / / 0.2] (neutral)4thvector [0.1 110.1 11 0.211 0.3110.3] (arousing / likeable)5thvector [0.0110.1 11 0.211 0.3110.4] (highly arousing / likeable)

[0077] Of course, it may be noted that in some cases, the value of a parameter may decrease to indicate arousal / likeability and in such cases, the selection of vectors may be reversed.

[0078] It may be noted that each of these vectors reflects probabilities that the user is experiencing a particular arousal / valence state. Thus, the values therein add up to 1. In this example, the states correspond to the 5 states of a SAM scale (-2, -1, 0, 1, 2), but other scales may apply in other examples and thus there may be more or fewer predetermined vectors in other examples.

[0079] In other examples, the neutral option (i.e. the third vector) may not be provided and instead the vector may be selected based on whether the response to the substance to be evaluated is more or less than one standard deviation away from the median, with the first or fifth vector being selected if the response to the substance to be evaluated is greater than one standard deviation away from the median, and the second or fourth vector being selected if the response to the substance to be evaluated is less than one standard deviation away from the median. In the event that the response to thesubstance to be evaluated is exactly the median, one of the second or fourth vectors may be selected at random or deterministically.

[0080] In some examples, there may be a plurality of parameters measured in relation to an ANS response. For example, as set out above, a plurality of factors may be extracted from an ECG and / or a skin response relating to different measured parameters such as heart rate and peak amplitude. In such examples, the vectors for each parameter may be combined, for example by providing an average thereof. In some examples, parameters which are more strongly associated with a particular response may have a higher weighting in the combination.

[0081] Figures 4A-D show example matrices / vectors which may be used in evaluating a substance.

[0082] Figure 4A is an example of a probability matrix which may be output by the neural response analysis module 102 as a first indication of a subject’s arousal and valence response to a particular substance, which in this example is an odorant, referred to herein as “fragrance X”. However similar principles may be applied when considering flavourings. In this example, the neural response analysis module 102 comprises a trained model, wherein the model had been trained using machine learning principles. In this example, the output uses a five-point response based on the responses used in SAM questionnaires. However, it will be appreciated that, where the system is configured based on a different measured response scale, the output may match that response scale, which might not be the same for both arousal and valence.

[0083] In this example, the response to “fragrance X” was acquired using a skullcap and 14 channels (acquired using 16 sensors, two of which provided reference values) of EEG data. The data was processed to extract a plurality of features to input to the trained model.

[0084] In a particular example, activity, mobility and complexity values (so called “Hjorth parameters”) for alpha, beta and gamma waves may be extracted. In one embodiment, all these features may be extracted for each channel, providing 3x3x14 = 126 features.

[0085] Alternatively or additionally, spectral entropy may be extracted e.g. using Welch’s method for alpha, beta and / or gamma waves (yielding in one example 3x14 = 42 features when spectral entropy is extracted for each of alpha, beta and gamma waves).

[0086] Further, alternatively or additionally, Energy and Entropy for wavelet decomposition at a plurality (e.g. 3) of levels may be extracted, using Welch’s method (Three levels, D1 , D2 and D3, giving 2x3x14=84 features).

[0087] Additional features may be provided based on Energy and Entropy for a first three IMF (intrinsic mode functions), for all channels (e.g. resulting in 2x3x14 = 84 features).

[0088] Differential alpha asymmetry may be calculated (by using pairs of interhemispherical channels) for different alpha features (e.g. Activity, Mobility, Complexity and Spectral Entropy). In this example, there may be 7 channel pairs and 4 alpha features (7x4 = 28 features).

[0089] Spectral slope for alpha, beta and gamma waves considering all channels may be determined (3 features)

[0090] Frontal alpha asymmetry, which is the difference between alpha power on the right hemisphere and alpha power on the left hemisphere (by using F3 and F4 channels), may contribute a further feature.

[0091] The ‘ratio of arousal’ between the beta power from frontal channels and alpha power from the same frontal channels (an example equation is set out below), may contribute a further feature.

[0092] Thus, in one example, 369 features may be determined from the 14 channels. In some examples, such a feature set (or a similar feature set, or a feature set comprising a subset of such features) may be provided to a suitably trained model. The model may be trained using training data comprising a plurality of training data entries acquired (e.g. measured) from training data test subjects. Each training data entry may comprise (i) a feature set corresponding to the feature set extracted from the EEG of a training data test subject while they experienced a flavouring / odorant, and (ii) questionnaire data indicative of the training data test subject’s reported response to the flavouring / odorant. The features extracted from the training data test subject’s EEG and used to train the model may be matched to the features to be extracted from the subject testing the particular substance to be evaluated, such that the same set of features is used in training the model and in use of the trained model. For example, this may comprise the 369 features described above, or a similar feature set, or a feature set comprising a subset of such features. Examples of training such a model are described in greater detail below.

[0093] In this example, the highest probability of the subject’s arousal and valence for fragrance X in the output probability matrix is associated with an arousal state of one (somewhat arousing) and valence state of one (somewhat likeable).

[0094] In this example, ECG and EDR data was obtained in order to assess arousal while the subject was also filmed to acquire facial expression data for use in assessing valence. In particular, ECG, EDR and video data was acquired while the subject was exposed to a set of odorants, including fragrance X. In total, the set included four odorants (although other examples could include a different number of odorants, including a significantly higher number). In this example, the ECG, EDR and video data relating to fragrance X was acquired at the same time as the EEG data referred to above. In addition, a resting state reading was acquired while the subject was not exposed to an odorant, but exposed to clean / odourless air.

[0095] The EDR data was analysed to extract: (i) a number of peaks relative to the resting state, (ii) an amplitude of peaks in the EDR relative to the resting state. The ECG data was analysed to extract: (iii) an average heart rate and (iv) an average R-R interval. A summary of the results for a particular subject is shown in Table 1 :Table 1

[0096] As can be seen from this result, the response to Fragrance X measured in the EDR was less than one standard deviation below the “number of peaks” median value and the “amplitude of peaks” median value. It may be noted that a greater number / amplitude peaks in an EDR response are generally indictive of high arousal. The response as measured in the ECG data was slightly less than one standard deviation below the average heart rate value and greater than one standard deviation above the Average R-R value. A higher R-R value indicates that each heart beat takes longer andis therefore indicative of a relaxed state, as is a lower heart rate. Based on these values, a vector may be selected which is indicative of a relaxed state.

[0097] Moreover, in examples herein, the parameters may be weighted differently from one another in determining a vector. In this example, it is noted that the amplitude of the peaks may be weighed lower than the other parameters.

[0098] The final arousal vector may be a weighted average of individual vectors selected for each parameter. The weights in this example were set as: 0.3 for Average HR; 0.3 for Average R-R Interval; 0.3 for Number of Peaks; 0.1 for Peaks Amplitude.

[0099] An average in this case may be a vector of:

[0100] [0.35 / / 0.29 / / 0.20 / / 0.11 / / 0.05],

[0101] This is rounded to [0.4 / / 0.3 / / 0.2 / / 0.1 / / 0.1], and since the probabilities do not add up to 1 , the lowest value of the vector is modified to arrive at [0.4 / / 0.3 11 0.2 / / 0.1 / / 0.0], i.e., the vector shown as Figure 4B.

[0102] As mentioned above, video data was also acquired and analysed to extract facial expression data. In particular, in this example, a count was made of a number of brow furrow events and a number of eyelid tightening events as the subject was exposed to each odorant. Without wishing to be bound by theory, both brow furrows and lid tightenings may generally be associated with a negative emotion. Similarly, nose wrinkles / inner brow raises may also be a negative emotion indicator, whereas a smile or smirk may indicate positive emotions.

[0103] Example data for a particular subject is summarised in Table 2:Table 2

[0104] Since fragrance X resulted in zero value for brow furrow and lid tightening, compared with the median value, this indicates that there was a “more positive” than the median emotion associated with fragrance X, leading to selection of the resulting positive ‘valence’ vector as shown in Figure 4C.

[0105] It may be noted that some facial landmarks, movements or states may be more difficult to assess clearly as being positive or negative (e.g., lip press). Moreover, as mentioned above, in other examples rather than indicating specific parameters associated with individual movements, some facial analysis software can extract information indicative of positive or negative valence, e.g:Negative— Anger, Disgust, Fear, Sadness, Confusion, ContemptPositive— Joy, Engagement,

[0106] Such analysis may be used in other examples to assess valence.

[0107] Having determined the arousal vector and / or the valence vector, these can be combined with the first indication of the subject’s arousal and valence.

[0108] For example, considering the probability matrix shown in Figure 4A, and the vectors shown respectively in Figure 4B and 4C, a mean of the values may be computed for each cell.

[0109] To consider a worked example, cell (-2,2) from the probability matrix shown in Figure 4A, indicative of a probability that the arousal level is 2 and the valence level is -2, is associated with a value of 0.0625, or approximately 0.06 as shown). The arousal vector value for that cell in Figure 4B (i.e. the cell for an arousal level of 2) has the value 0.0 and the valence vector has the value 0.1 in Figure 4C (i.e. the cell for a valence level of -2). The final value in M2 for cell / position (-2,2) is calculated as the mean of the three aforementioned values (= 0.0542, or approx. 0.05). A weighted average may be used in other examples.

[0110] A result of combining the matrix and the vectors is shown in Figure 4D. As can be seen from this matrix, the most likely value for arousal and valence is no longer [valence = 1 , arousal = 1], but is instead indicated as [valence = 1 , arousal = -2], indicating that the odorant is more relaxing than originally indicated.

[0111] Moreover, in this particular example, the odorant Fragrance X had previously been evaluated using a SAM questionnaire and was associated with an average arousal and valence of [valence = 2, arousal = -2], Thus, in this case, the accuracy of the interpretation of the subject’s response is likely to have been increased.

[0112] While in this example a probability matrix was determined and modified, this need not be the case in all examples. In some examples, an estimation of a most likely valence and / or arousal state may be produced from neural response data and modified with estimations of valence as determined from expression analysis and / or arousal asdetermined from ANS response. For example, an average (which may be a weighted average) may be determined.

[0113] Without wishing to be bound by theory, it is noted that the EEG response is impacted by both valence and arousal. Thus, while the signal is highly valuable as it relates to data indicative of a subject’s central nervous system (i.e. the brain and the spinal cord, wherein the brain is responsible for receiving, processing and responding to sensory data), the interpretation of the signal is somewhat complex. However, by providing data which is separately indicative of arousal and / or valence, this result can be disambiguated.

[0114] It may be noted that facial expression can also be somewhat subjective depending on a subject’s tendency to express themselves visually. In some cases, a subject may be encouraged to allow their face to reflect their emotion. However, as in some examples herein the facial expression data is acquired at the same time as the neural response data, this may help overcome any ambiguity in the facial expression data. Similarly, any atypical ANS response may be at least partially resolved by neural response data.

[0115] Moreover, examples herein may utilise a resting state and / or a comparison to an average state. This may usefully establish an individual subject’s response range such that responses from subjects with different response ranges may be somewhat normalised.

[0116] In examples herein, fragrance X may be tested by a plurality of subjects, who may be untrained. In this way, the responses may be more natural and / or more closely reflect the responses of the intended average consumer of goods comprising fragrance X. An average or combined response to fragrance X may be acquired from, for example at least 10, or at least 15 or at least 20 subjects, and it may be determined whether fragrance X sufficiently closely matches a target classification. For example, a difference (which may be a numerical difference) between an evaluation output and a target evaluation value may be determined, and it may be determined if this is less than a threshold value (which may be a different threshold for valence and for arousal). If so, fragrance X may be selected or recommended for use in a product. For example, it may replace an ingredient in an existing composition or may be associated with a product which is intended to create a particular response in the user.

[0117] Figure 5 shows an example of processing circuitry, which may comprise one or more processors, in association with a machine readable medium. The machine- readable medium stores instructions which, when executed, cause the processing circuitry to carry out tasks.

[0118] In particular, the instructions may comprise instructions for performing evaluation of a substance comprising an odorant or a flavouring. The instructions comprise instructions to obtain neural response data of a subject’s response to a substance. This may comprise receiving neural response data, e.g. an EEG, acquired from a skullcap or the like carrying electrodes which are used to directly monitor the subject. In other examples, it may comprise receiving data indicative of the subject’s neural response. In some cases, this data may be derived from an EEG or the like.

[0119] The instructions further comprise instructions to cause the processing circuitry 500 to determine a first indication of the subject’s arousal and / or valence response to the substance based on the neural response data, for example based on a predetermined relationship. For example, the first indication may comprise a probability matrix as shown in Figure 4A. In other examples, the first indication comprises at least one of a probability matrix / vector associated with a valence response and a probability matrix / vector associated with an arousal response. In other examples, the first indication may comprise an evaluation such as an estimated valence and / or arousal of the subject. In some examples, the first indication of the arousal and / or valence is determined using at least one model trained using machine learning techniques. In some examples, two models may be trained, one in relation to valence and one in relation to arousal. In some such examples, the output of such models may be combined to provide a combined indication of valence and arousal.

[0120] The instructions further comprise instructions to cause the processing circuitry 500 to obtain (for example determine) (i) an arousal response profile for the subject based on autonomic nervous system biometric data obtained in response to the substance to be evaluated, and / or (ii) a valence response profile for the subject based on facial analysis of images acquired while the subject was exposed to the substance. For example, these may comprise an estimation of the subject’s valence and / or arousal response acquired separately. Biometric data, which may be associated with an ANS response, may be used as the basis for evaluating the subject’s arousal response and / orfacial expression analysis may be used as the basis for evaluating the subject’s valence response.

[0121] As set out above, in some examples, obtaining the response profiles may comprise receiving at least one dataset indicative of a subject’s arousal and / or valence response to each of a set of substances (including the substance to be evaluated) and determining the response profile based on the position of the data indicative of the subject’s response to the substance being evaluated within the dataset. For example, an average response from the dataset and a measure of variability in the dataset may be used to select one of a set of predetermined response profiles based on the average and the variability. This may normalise the degree of response from different subjects such that subjects who have an extreme reaction to substances do not unduly influence evaluation of the substance under test.

[0122] In some examples, a plurality of valence response profiles relating to different parameters indicative of valence may be combined to provide the valence response profile. Similarly, a plurality of arousal response profiles relating to different parameters indicative of arousal may be combined to provide the arousal response profile. For example, this may comprise determining an average, which may comprise a weighted average.

[0123] The instructions further comprise instructions to cause the processing circuitry 500 to modify the first indication of the subject’s arousal and / or valence response to the substance based on the arousal response profile and / or the valence response profile to obtain a second indication of the subject’s arousal and / or valence response. For example, this may comprise increasing or decreasing an initially estimated arousal and / or valence response. In some examples, this comprises combining the first indication and the arousal and / or valence response profiles.

[0124] The instructions further comprise instructions to cause the processing circuitry 500 to evaluate the substance based on the second indication. The evaluation may for example comprise an indication of arousal and / or valence of the subject’s response to the substance and / or a difference between an arousal and / or valence of the subject’s response to the substance and a target arousal and / or valence response.

[0125] In some examples, the instructions may comprise instructions to cause the processing circuitry 500 to train a model using neural response data and survey data indicative of the response of a plurality of subjects to a plurality of different substances,wherein the model may be used to provide the first indication. For example, training the model may utilise machine learning principles. In some examples, the model is trained using K-Nearest Neighbours (KNN) methods. In some examples, the model is a probability, or probabilistic, classifier.

[0126] In some examples, the instructions may be used to cause the processing circuitry to evaluate the responses of each of a plurality of users.

[0127] In some examples, the instructions may further comprise instructions to cause the processing circuitry to obtain neural response data, arousal response data and / or valence response data to an odorant and / or flavouring. For example, the instructions may cause the processing circuitry to control apparatus such as the apparatus shown in Figure 6.

[0128] Figure 6 shows an example of a system comprising EEG apparatus 602 (in this example a skullcap), ANS sensing apparatus 604 (in this example a finger electrode to measure a skin conductivity, although other apparatus may be used in other examples) and an image capture device 606, in this example a webcam. The EEG apparatus 602, ANS sensing apparatus 604 and image capture device 606 are configured to provide readings to processing circuitry 100 which in this example is processing circuitry as shown in Figure 1. In this example, the system further comprises an olfactometer 608 comprising a control unit to manage air flow given to an olfactory outlet. An air pump may be used to generate an air flow which is then mixed with air carrying the odorant to be evaluated.

[0129] Similar apparatus, absent the olfactometer 608, may be provided to measure a subject’s response to a flavouring. As noted above, in some examples, standardisation may be carried out such that each flavouring is delivered in a consistent way (e.g. as a particular quantity of liquid and / or from a similar vessel). In some examples steps may be taken to reduce impact of texture and / or appearance of the flavouring on the subject’s response. For example, controlled lighting conditions and / or colours of delivery vessels may be used to minimise visual differences between flavourings. A delivery system may be selected which minimises chewing. This may reduce the impact of texture and / or of facial movements which may make the facial analysis more difficult to interpret and which may add artifacts to neural response data. In such cases, a liquid or dissolvable delivery system may be used.EXAMPLES OF TRAINING A MODEL

[0130] Examples herein may utilise a model which is derived using machine learning principles. The model may be used for example by the neural response analysis module 102 to derive the first indication of a subject’s arousal and / or valence response to a substance. Examples of training such a model are now described in greater detail.

[0131] Interpretation of EEG data to extract an arousal and / or valence response using machine learning has been discussed, for example in “Predicting Exact Valence and Arousal Values from EEG” by Galvao et al (Sensors 2021 , 21 , 3414), incorporated herein by reference. In that paper, various methods of extracting Valence and Arousal are discussed. The authors describe using publicly available databases (AMIGO, DEAP and DREAMER) made up of SAM questionnaire responses to videos as well as various biometric indicators of arousal and / or valence, of which the EEG signal was utilised. As the authors intended to identify continuous valence and arousal values, they used regression-based machine learning methods, concluding that random forest and K- nearest neighbour (KNN, K = 1) methods performed best, with KNN achieving the best result overall.

[0132] Techniques discussed in this document may be utilised in examples described herein.

[0133] In a particular example, for evaluating responses to odorants, a plurality of tests were carried out in which, in each test, EEG data was collected for a single odorant tested by a single subject using a 14-channel headset that describes the following positions: AF3, F7, F3, FC5, T7, P7, 01 , 02, P8, T8, FC6, F4, F8, and AF4. Other channels may be used in other examples. The same channels used to develop the model may be used by a test subject to evaluate a substance under test. The data was analysed to extract features therefrom. The features may comprise the features as described above for each test subject (e.g. the feature set comprising 369 features described for the test subjects). However, a subset of these features, or other features, may be used in other examples.

[0134] In an example, the training data comprised features indicative of:A: Hjorth parameters (specifically Activity, Mobility and Complexity for each of the alpha, beta and gamma waves).B: Spectral entropy i.e. a measure of entropy extracted from the distribution of frequencies for a signal (spectral analysis) to estimate the uncertainty and randomnessof the distribution of frequencies from the signal. This may be calculated by determining the spectrum of the signal and deriving the power spectral density (PSD). In examples, the PSD is normalised to provide a distribution of frequencies summing up to 1. An entropy formula may be applied thereto using equation (x log(x)). In addition to spectral entropy, spectral energy may be determined from the wavelet decomposition for alpha, beta and / or gamma waves. As noted above, spectral entropy may be extracted for alpha, beta and / or gamma waves. In addition, energy and entropy for wavelet decomposition at a plurality (e.g. 3) of levels and for a plurality (e.g. 3) of intrinsic mode functions may be extracted for each channel.C: Differential Alpha Asymmetry may be determined by considering interhemispherical channel pairs and alpha features for activity, mobility, complexity and spectral entropy.D: Spectral slope. The PSD may be plotted and divided into ranges of frequencies (alpha - 8-12 Hz; beta - 12-30 Hz; gamma - 30-45 Hz). For each segment of the signal, a linear line is fitted, and the slope extracted. It is noted that the PSD of normal EEG signals generally has a peak for low frequencies and then a slope down towards higher frequencies, which may be called the signal's “drop-off’. The slope of the “drop-off” appears to be related to arousal level.E: Frontal alpha asymmetry. This is the symmetry between alpha power on the left and on the right side of the frontal brain. The PSD for left (F3 channel) and right (F4 channel) of the brain may be determined and an equation that evaluates asymmetry (e.g. FAA = log(F4_power / F3_power) may be applied.F: A “ratio of arousal” between channels defined as:

[0135] In addition, SAM survey data was acquired from the subject immediately after the EEG was acquired with respect to an odour. The survey data included:G: a survey response value for valence each associated with one of N values (e.g. -2, -1, 0, 1 and 2); andH: a survey response value for arousal, each associated with one of the N values.

[0136] While steps may be taken to exclude subjects who have anosmia and / or hyperosmia, (or related conditions which impact the sense of smell), and in someexamples people who use their sense of smell for a living, the subjects may be taken from the general population. Similar steps may be taken when evaluating a flavouring. In such cases, any or all of people who use their sense of taste for a living, people with ageusia and / or so-called ‘supertasters’ may be excluded. In other examples, the subjects may be taken from a population made up of a target market for a product which may be associated with the odorant / flavouring. In some examples, similar criteria may be applied when evaluating a particular substance such as Fragrance X described above.

[0137] In a particular example, the training set had 1067 entries acquired from 41 subjects balanced between men and women who had been exposed to 6 different odours. The odours included both invigorating and relaxing fragrances, as well as some that were considered as neutral. Each subject's EEG data for each odour was divided into 4 second time segments such that each test provided a plurality of data samples.

[0138] In some examples, the training data may comprise or include data indicative of a response to a stimulus other than odour / taste. For example, the training data set may include data taken in response to a visual stimulus (e.g. an image or a video)

[0139] Such data may be more readily obtainable than data in relation to odorants.

[0140] In some examples, a KNN classification algorithm may be used. While other algorithms may be used in other examples, one example of such an algorithm is based on a module from the Scikit-learn machine learning library for the Python programming language. For example, the sklearn. neighbors. KNeighborsClassifier module may be used with the number of neighbours set to 12 and a "Manhattan" distance metric, although other variables may be used in other examples

[0141] Two models were trained separately, one relating the set of extracted features to the SAM responses for arousal, and a second relating the set of extracted features to the SAM responses for valence.

[0142] In use of the models as a classifier, the features extracted from the EEG of a test subject were evaluated separately to determine a valence and arousal value. As described above, a probability value may be output for each individual valence and arousal state. As will be familiar to the skilled person, this may utilise the sklearn. neighbors. KNeighborsClassifier.predict_proba module within sklearn. neighbors. KNeighborsClassifier module to return a probability array having a length equal to the number of valence / arousal states (e.g. as in examples herein, thereare five such states). In some examples, a combined matrix was determined by averaging the probability values of the corresponding cells.

[0143] The trained classifier was evaluated using a retained portion of the data set comprising 107 entries. This resulted in a mean absolute error (MAE) of 0.76 and 0.78 for arousal and valence, respectively.

[0144] In a further evaluation example, the model was evaluated using responses to Fragrance X as detailed above. A mean absolute error was determined by comparing the output of the model in relation to 28 neural data samples to SAM data associated with Fragrance X. This gave a mean absolute error of 2.29 for Arousal and 0.57 for Valence. Thus it could be concluded that the model was somewhat better at evaluating valence than arousal.

[0145] However, the methods set out above were then applied, and determined arousal response profiles and valence response profiles (e.g. as shown in Figures 4B and 4C) were used to modify the first indications, which in this example were probability matrices (e.g. based on the same measured variables set out in tables 1 and 2 and for example as shown in Figure 4A). This results in improved mean absolute error of 1.64 for Arousal and 0.18 for Valence. Thus, applying the arousal response profiles and valence response profiles significantly increased the accuracy of the prediction.

[0146] It should be recognized that features and aspects of the various examples provided above can be combined into further examples that also fall within the scope of the present disclosure, and / or the skilled person will be aware of alternative features which fall within the scope of the claims. The scope of the claims should not be limited by the above examples but should be given the broadest interpretation consistent with the description as a whole.

Claims

CLAIMS1. A method of evaluating a substance comprising an odorant and / or a flavouring comprising: obtaining neural response data of a subject’s response to the substance; determining a first indication of the subject’s arousal and / or valence response to the substance based on the neural response data; obtaining at least one of: i. an arousal response profile for the subject based on autonomic nervous system biometric data obtained in response to the substance to be evaluated, and ii. a valence response profile for the subject based on facial analysis of images acquired while the subject was exposed to the substance; modifying the first indication of the subject’s arousal and / or valence response to the substance based on the arousal response profile and / or the valence response profile to obtain a second indication of the subject’s arousal and / or valence response wherein: if the first indication is an indication of the subject’s arousal response, modifying the first indication comprises modifying the first indication based on the arousal response profile; if the first indication is an indication of the subject’s valence response, modifying the first indication comprises modifying the first indication based on the valence response profile; and if the first indication is an indication of the subject’s arousal and valence response, modifying the first indication comprises modifying the first indication based on at least one of the arousal response profile and the valence response profile; the method further comprising: evaluating the substance based on the second indication.

2. A method according to claim 1 wherein: the first indication of the subject’s arousal and / or valence response to the substance comprises a probability matrix or vector associating a probability with each of a plurality of arousal response levels and / or valence response levels.

3. A method according to claim 1 or claim 2 comprising: receiving at least one dataset indicative of a subject’s arousal response to each of a set of substances, the set of substances comprising the substance to be evaluated; and wherein determining the arousal response profile to the substance to be evaluated comprises determining the arousal response profile based on the position of the data indicative of the subject’s arousal response to the substance to be evaluated within the dataset.

4. A method according to any preceding claim comprising: receiving at least one dataset indicative of a subject’s valence response to each of a set of substances, the set of substances comprising the substance to be evaluated; and wherein determining the valence response profile to the substance to be evaluated comprises determining the valence response profile based on the position of the data indicative of the subject’s valence response to the substance to be evaluated within the dataset.

5. A method according to claim 3 or claim 4 wherein determining the response profile(s) comprises determining an average response from the dataset and a measure of variability in the dataset; and selecting a predetermined response profile based on the average and the variability.

6. A method according to claim 5 comprising acquiring a plurality of data sets, each dataset relating to a different parameter indicative of a subject’s valence and / or arousal response; determining a valence response profile or an arousal response profile for each parameter; and combining the valence response profiles or the arousal response profiles for each of a plurality of parameters to determine a valence response profile or arousal response profile.

7. A method according to any preceding claim comprising receiving electroencephalography, EEG, data, the data comprising data indicative of a plurality of EEG channels; processing the EEG data to extract the neural response data therefrom, the neural response data comprising at least one of: at least one Hjorth parameter; an indication of spectral entropy; an indication of energy, an indication of frontal alpha asymmetry; an indication of spectral slope; an indication of differential alpha asymmetry; and a ratio between channels.

8. A method according to any preceding claim wherein the autonomic nervous system biometric data comprises any or any combination of:ECG data; pulse monitoring data; dermal response data; data indicative of respiration; data indicative of blood oxygen level; data indicative of skin temperature; data indicative of pupil size or response; and data indicative of gut activity.

9. A method according to any preceding claim wherein the determining the first indication of the arousal and / or valence response comprises determining the first indication of the arousal and / or valence response using at least one model trained using machine learning techniques.

10. A method according to claim 9 further comprising training the model(s) using neural response data and survey data indicative of the response of a plurality of subjects to a plurality of different substances.

11. A method according to any preceding claim wherein the evaluation comprises an indication of arousal and / or valence of the subject’s response to the substance.

12. A method according to claim 11 wherein the evaluation comprises a difference between an arousal and / or valence of the subject’s response to the substance and a target arousal and / or valence response.

13. A method according to any preceding claim comprising obtaining neural response data, arousal response data and / or valence response data to the substance from the subject.

14. A method according to any preceding claim, wherein the method is carried out for each of a plurality of subjects, and the substance is evaluated based on a combination of second indications of the subjects’ arousal and / or valence responses.

15. A method according to any preceding claim wherein the substance comprises an odorant.

16. Apparatus comprising processing circuitry comprising: a neural response analysis module configured to: obtain neural response data of a subject’s response to a substance comprising an odorant and / or a flavouring; and determine a first indication of the subject’s arousal and / or valence response to the substance based on the neural response data; a response modification module configured to: obtain at least one of: an arousal response profile for the subject based on autonomic nervous system biometric data obtained in response to the substance to be evaluated; and a valence response profile for the subject based on facial analysis of images acquired while the subject was exposed to the substance; wherein the response modification module is further configured to:modify the first indication of the subject’s arousal and / or valence response to the substance based on at least one of the arousal response profile and the valence response profile to obtain a second indication of the subject’s arousal and / or valence response wherein: if the first indication is an indication of the subject’s arousal response, modifying the first indication comprises modifying the first indication based on the arousal response profile; if the first indication is an indication of the subject’s valence response, modifying the first indication comprises modifying the first indication based on the valence response profile; and if the first indication is an indication of the subject’s arousal and valence response, modifying the first indication comprises modifying the first indication based on at least one of the arousal response profile and the valence response profile; and an evaluation module configured to evaluate the substance based on the second indication.

17. Apparatus according to claim 16 further comprising at least one of: a neural response monitoring apparatus configured to acquire neural response data and provide the neural response data to the processing circuitry;ANS response monitoring apparatus configured to acquire autonomic nervous system biometric data and provide the autonomic nervous system biometric data to the processing circuitry; and an image capture device configured to acquire images of the subject, and to provide the images to the processing circuitry.

18. A non-transitory machine-readable medium storing instructions wherein the instructions, when executed, cause processing circuitry to carry out the method of any of claims 1 - 15.

Images