System and method for detecting foreign substances in drinks

Abstract

A wearable sensing device (120) and associated method are provided for determining whether or not a drink (112) has been spiked. The sensing device (120) includes a sensor (310) for holding against a container (110) including the drink (112). The sensor (310) includes at least first and second light sources (170) which are configured to be placed against the side of the container (110) to transmit light at first and second wavelengths into the container (110), wherein the first wavelength is different from the second wavelength. The sensor (310) further includes at least one photoreceptor (160) configured to be placed against the side of the container (110) to receive light which has been transmitted into the container (110) by the light sources (170) and reflected out of the container (110). The intensity of the light received by the photoreceptor (160) for the first and second wavelengths is compared with reference characteristics to determine whether or not the drink (112) has been spiked.

Description

[0001] SYSTEM AND METHOD FOR DETECTING FOREIGN SUBSTANCES IN DRINKS

[0002] Field

[0003] The present application relates to a system and method for sensing drinks, for example to determine whether or not a drink has been spiked.

[0004] Background

[0005] Drink spiking is the deliberate act of adding a substance to a drink. Typically drink spiking is perpetrated by someone intending to make a victim more vulnerable to assault. It is an important social issue to overcome as it can cause victims severe danger, anxiety, and physical and / or mental harm [2, 15]. At the same time, methods for protecting people from spiked drinks need to be unobtrusive so they are acceptable in social contexts. A study among student populations has shown that almost 8% of students have had their drink spiked at some point

[0016] , highlighting how drink spiking is a significant threat to people. The public perception is that drink spiking is limited to slipping drugs or sedatives into alcoholic drinks, but drink spiking can also target non-alcoholic drinks, such as water, soda, soft drinks, and juices.

[0006] Detecting drink spiking is unfortunately difficult as the spiking typically happens in crowded social occasions where it is difficult to keep a consistently watchful eye on a drink container and its contents. Substances that are commonly used in drink spiking also are not readily visible, making it hard to realize when a drink has been tampered with. Currently, the primary countermeasure has been to design guidelines and to provide advice for raising awareness of the problem. Such advice may include, for example, requesting sealed products for purchase, keeping a watchful eye on a drink, avoiding leaving drinks unattended and so on. Unfortunately, these methods have limited efficacy as people tend to underestimate the possibility of spiking. Further, spiking may occur too quickly for the victim to realise what is happening, or else the victim may be distracted at the moment of spiking.

[0007] Some technological solutions have been developed for identifying spiked drinks. In particular, drink spiking has been investigated mainly by analyzing the chemical composition of liquids. For instance, a method using rapid capillary zone electrophoresis was proposed to identify benzodiazepine drugs (aka benzos) in (spiked) beverages that include Coca-Cola, orange juice, beer, bourbon, and Bacardi

[0017] . Spectroscopy also has been used to identify spiked drinks. For example, a fluorescence spectroscopy method was investigated to detect flunitrazepam (Rohypnol benzodiazepine) in colourless alcoholic liquids such as vodka and tequila

[0011] . However, such solutions tend not to be well-suited for everyday interactions, for example, they may require sophisticated equipment such as an apparatus for performing chromatographic tests. Such an apparatus may, for example, be expensive and have limited portability.

[0008] Other methods have also studied the use of ultraviolet and electrochemical approaches for detecting benzodiazepines in liquids

[0013] . While these works demonstrated that spiked drinks could be detected, they again generally involve specialized and expensive instruments which cannot be easily integrated into wearables or deployed at large-scale. In parallel to this, industrial manufacturers, such as DrinkSavvy, have designed specialized cups, glasses and straws that instantly change the colour of a drink as it gets spiked. While this approach is convenient for continuous monitoring, it may be avoided by using similar utensils. Moreover, this product only works for a limited set of drugs. The following documents give a brief overview of some other activity in this field. US2017242045 discloses methods and apparatus for detecting compounds in liquids. The apparatus is not an electronic device; rather it is a kind of artificial fingernail that can be attached to nails. EP1332720A1 discloses an instrument for measuring concentration. This instrument is intended to measure blood sugar concentration by clamping the instrument on the earlobe and using light reflectivity. A method to screen for flunitrazepam in vodka and tequila using fluorescence spectroscopy by Leesakul et al is presented in Luminescence 2013 Jan / Feb, 28(1 ), 76-83 (see https: / / pubmed.ncbi.nlm.nih.gov / 22354877 / ). This approach relies on a spectrometer, so that it is generally not portable and is generally operated by an expert. US11448587B1 discloses devices and methods for detecting drugs, such as gamma hydroxybutyrate and gamma butyrolactone, using infrared spectrometry. US2016131577 discloses a method for the contactless detection of psychoactive components in a liquid and an apparatus therefor, by emitting a substantially monochromatic light at least at two different wavelengths and detecting the reflection in a free surface of the liquid by a photodetector. US2016146726 is directed to wearables having a spectrometer for analysing a chemical composition of a substance. Although previous inventions have utilized spectroscopy-based analysis for drug detection, using light in everyday settings and pervasive interactions presents significant challenges. Factors such as motion, varying luminosity, and data quality all complicate the effectiveness of light-based detection methods.

[0009] There remains interest in the development of further devices and strategies to protect against drink spiking in a convenient and effective manner.

[0010] Summary

[0011] The invention is defined in the appended claims.

[0012] The present application relates to a system and method for sensing drinks, for example to determine whether or not a drink has been spiked.

[0013] A wearable sensing device and associated method are provided for determining whether or not a drink has been spiked. The sensing device includes a sensor for holding against a container including the drink. The sensor includes at least first and second light sources which are configured to be placed against the side of the container to transmit light at first and second wavelengths into the container, wherein the first wavelength is different from the second wavelength. The sensor further includes at least one photoreceptor configured to be placed against the side of the container to receive light which has been transmitted into the container by the light sources and reflected out of the container. The intensity of the light received by the photoreceptor for the first and second wavelengths is compared with reference characteristics to determine whether or not the drink has been spiked.

[0014] Brief Description of the Drawings

[0015] Embodiments of the invention are described below, by way of example only, with reference to the accompanying drawings, in which:

[0016] Figure 1 is a schematic diagram showing an overview of various aspects of an approach described herein to detect whether a drink has been spiked, including a smart ring, a set of drinks, and a set of potential contaminants. Figure 1 A is a set of images showing glasses of milk to which an increasing amount of cocoa powder have been added.

[0017] Figure 2 is a schematic diagram showing an example of a sensing system as described herein to detect if a drink has been spiked.

[0018] Figure 3 is a schematic diagram showing an example of a smart ring such as may be used in the sensing system of Figure 2.

[0019] Figure 4 depicts on the left-hand side a schematic diagram illustrating various positions for wearing a smart ring such as shown in Figure 2, and on the right-hand side an example of measuring the hand-grip strength of a user to help calibrate the sensing system shown in Figure 2.

[0020] Figure 5 is a schematic diagram which provides an example of the overall sensing system and process for assessing whether or not a drink has been spiked as described herein.

[0021] Figure 6 provides experimental data arranged in two rows of graphs. Each graph in the top row shows the sensed light values for a variety of different drinks, whereas each graph in the bottom row shows the sensed light values for milk containing an increasing amount of cocoa powder. Each row comprises three plots corresponding to three different finger positions for wearing the smart ring, namely position 1 (fifth finger), position 2 (third finger) and position 3 (second finger), corresponding to plots (a) and (d); (b) and (e); and (c) and (f) respectively.

[0022] Figure 7 provides 5 plots in sections (a)-(e) of experimental results, wherein each plot shows the measured light value produced in certain circumstances. Section (a) shows results for water having various contaminants and various smart ring positions. Section (b) shows results for water have various contaminants and subject to shaking. Section (c) shows results for various ambient lighting. Sections (d) and (e) show results for water having various contaminants using a different sensor from that used in sections (a)-(c). In section (d) the sensor is placed to the side of the glass container, in section (e) the sensor is placed underneath the glass container.

[0023] Figure 8 provides 6 plots in sections (a)-(f) of experimental results showing the variation in measured light values according to factors such as light / dark beer, ambient lighting conditions, and cup type.

[0024] Figure 9 presents Table 1 showing the performance of drink spiking classification using machine learning for different light conditions and features.

[0025] Figure 10 is a high-level flowchart which provides an example of using the sensing system of Figure 2 as described herein.

[0026] Figure 11 (a) to (d) shows the stages in a processing pipeline of an embodiment.

[0027] Figure 12 shows how reflected light levels vary in dependence on motion and background light.

[0028] Figure 13 shows the processing steps involved in removing noise due to movement and background light.

[0029] Detailed Description

[0030] General

[0031] People increasingly carry wearable devices (wearables) and the capabilities of these devices have grown so that it is increasingly viable to harness such devices to support everyday interactions. As described herein, a new use of wearables is to safeguard against drink spiking, the deliberate act of adding substances to another person’s drink. A pervasive sensing approach has been developed that re-purposes the optical sensors in off-the shelf wearables to identify spiked drinks by analysing differences in light reflectivity resulting from small particles inside the drink. This approach provides a wearable device based partly on some aspects of a smart ring design. It is therefore seen how pervasive sensing enables innovative applications and how smart wearables can be re-purposed to support personal safety.

[0032] Light reflectivity methods (e.g. in the green spectrum) have mostly been used in photoplethysmography to estimate heart rate via propagation of light through the body [3]. Several works have investigated approaches to re-purpose sensors from smart devices to identify liquids and materials, such as alcoholic drinks, sugar, liquid density, liquid surface tension and liquid viscosity [5, 7-9, 19, 21]. A new sensing method and smart ring wearable to identify drink spiking are described herein. Such a smart ring may be equipped with an optical sensor to identify changes in the light reflectance from a drink and hence to detect when the drink has been spiked. In broad terms, the smart ring can be considered as an electronic wearable device. In operation, the smart ring may be positioned against a glass (in particular, an outer surface of the glass) and provide a light source to shine light into the container. The smart ring further includes at least one light receptor in one or more portions of the spectrum (e.g. red light and blue light) to measure reflected light. It has been found that the light reflectivity of a liquid in the glass changes if the drink has been spiked. The measurements from the light receptor(s) may therefore be analysed to determine whether or not a drink in the glass has been spiked. The smart ring may communicate with a smartphone app or other similar facility to alert a user if their drink has been spiked. This approach is easy to use and can be readily linked to a smartphone, a smartwatch, or similar device. Accordingly, the smart ring may benefit anybody seeking to protect themselves from drink spiking.

[0033] The smart ring differs from many other devices which address the issue of spiked drinks in that it is an electronic wearable device (smart ring) which includes and uses light sensors (sources and receptors) in different wavelength ranges (colours) of the spectrum, such as red, green and blue. The smart ring may communicate with a smartphone app to indicate whether or not a drink has been spiked. The smart ring is reusable, does not rely on extra chemicals, and does not need to be submerged in the liquid to determine the presence of a foreign substance in the drink (spiking). This is in contrast to some existing devices which rely on ‘wet’ chemical reactions to determine whether or not a drink has been spiked. Such existing devices typically involve submerging a test material physically into the liquid. Not only is this approach more complicated for a user, but also the chemicals represent a consumable which generally becomes depleted with use. In many cases, such existing devices are not reusable - i.e. they may be used only once to detect spiking; in other cases the devices may only operate a limited number of times before the chemical reactants used to detect spiking are exhausted.

[0034] The smart ring disclosed herein has some high-level similarities to certain existing devices that are intended to measure human health indicators, such as blood oxygenation. However, these health devices have a light source directed inwardly (towards the centre of the ring) so that the light falls on skin of the digit on which the ring is located. In contrast, the smart ring disclosed herein has the light source directed radially outwards to support testing drinks for spiking. Overview

[0035] Drink spiking occurs when a drug, sedative or other such substance (or multiple such substances) are put into the drink without the knowledge or permission of the person that the drink belongs to. As described herein, a sensing system including a sensing device has been developed to enhance the personal safety of individuals in relation to drink spiking. This system is referred to herein as the Hedgehog system (having regard to the spikes of a hedgehog). The sensing system re-purposes light sensors, which are often utilised in various existing wearable devices, in order to identify spiked drinks. As described herein, the performance of the sensing solution to identify spiked drinks under different ambient conditions and real- life situations has been experimentally investigated and confirmed. In particular, results presented herein indicate that the Hedgehog system may have an accuracy of up to nearly 90% for identifying spiked drinks. (As discussed below, the outcome may be somewhat dependent on circumstances such as the type of drinks that have been spiked, the substances used for spiking, the glass or other container used for holding the drink, the general lighting conditions, and so on).

[0036] The sensing system described herein may be based on a ‘smart’ ring which typically integrates various functionality, such as a light sensor and / or communications support. In an existing smart ring, a sensing system uses one or more light sensors to take measurements from the skin of individuals. In contrast, the approach described herein for a Hedgehog sensing system externalizes the light sensor to take measurements from drinks as the user interacts with the drinks. In other words, the light sensor operates in a direction radially outwards from the ring.

[0037] By way of example, a light sensor is placed on the outer surface of a liquid container, such as the side of a glass. The container (vessel) is generally transparent or translucent to allow the light sensor to detect light that has passed through the liquid in the vessel (and through the container itself). Light scattering and reflectivity principles are then applied to analyze the measurements of the received light. Further details above the light scattering and reflectivity principles can be found, inter alia, in

[0022] .

[0038] (In general, light refraction typically occurs when light passes from one optical medium into another optical medium. Light reflection typically occurs when light encounters a smooth surface which is not fully transparent or which results in total internal reflection. Light scattering typically occurs when light encounters a rough surface or particles suspended in a fluid. In the present application, for conciseness we will generally refer to reflected or scattered light, but it will be understood that such light may typically experience reflection, refraction, or scattering or any combination thereof.

[0039] In one example, a smart ring prototype includes an embedded light source and one or more embedded light receptor(s), e.g. a laser diode and a photoresistor respectively, to act as a sensor. The photoresistor measures the intensity of light reflected back from the beam produced by the laser diode. The ring further includes electronic support to connect to a computing board which analyses the measured values of the reflected light. The example implementation utilises a wireless M5StickC PLUS ESP32 development board (see https: / / docs.m5stack.com / en / core / m5stickc_plus) to control the sampling frequency of the light sensor and to upload the samples (sensed light values) to a server in real time. It will be appreciated that other development boards which provide analogous functionality to the M5StickC PLUS ESP32 development (support) board may also be used. In other implementations, the development board may be replaced, for example, by an app running on a smartphone or similar device. Figure 1 is a schematic diagram showing an overview of various aspects of an approach described herein to detect whether a drink has been spiked. In particular, Figure 1 comprises four main sections: (a) a smart ring and associated board; (b) characterisation of the drinks; (c) spiking the drinks with one or more medical compounds; and (d) spiking the drinks with one or more soluble pills in different types of drinks and different types of drinking vessels.

[0040] With reference to section Figure 1 (a), this shows a user hand and wrist from four different angles. The user is holding a glass 110 which includes a liquid such as dark beer. The user is wearing a smart ring 120 as described herein and also has a support board 130 attached to their wrist by a wrist band. The smart ring 120 and the support board 130 are in wireless communication with one another, and the support board 130 may be further in contact with a remote server (not shown in Figure 1 ). In a commercial implementation, the functionality of support board may be incorporated into a smart device such as a smartphone, for example in the form of a suitable app. Since a user will generally already take their smartphone with them to a pub or bar, this smartphone configuration dispenses with the user having to take an additional device such as a support board 130 for operating with the smart ring 120.

[0041] With reference to section Figure 1 (b), this illustrates the following drinks: (A) water, (B) carbonated soft drink (cola), (C) beer, (D) milk, (E) cold tea, and (F) grape juice. Each drink is in a glass (a transparent container). These six drinks are commonly available and widely consumed and are susceptible to manipulation (spiking) in a social context.

[0042] A calibration procedure is performed with respect to the above drinks in their proper form - i.e. without spiking or any other contamination by a foreign substance. In particular, the light measurements obtained using the light source and light sensor are analysed to develop ‘fingerprints’ (characteristics) that serve as a reference for each type of drink in the absence of spiking. This then allows subsequent light measurements for a given type of drink to be compared with the fingerprints for that type of drink. If there is a match to the reference fingerprints, this provides an indication that the drink is uncontaminated. However, if there is a discrepancy or deviation, such that the light measurements do not match the reference fingerprints, this provides an indication that the drink may now be contaminated (spiked).

[0043] With reference to section Figure 1 (c), this shows three glasses S1 , S2, S3 containing water. Each glass has been spiked (contaminated) with a particular medical compound: S1 includes a 500 mg tablet of aspirin; S2 includes a 500 mg tablet of paracetamol; and S3 includes a 30 mg tablet of mirtazapine. In general terms, the aspirin and paracetamol dissolve at least partly into the water, so are less visible as a contaminant, while the mirtazapine leaves the water with a cloudier appearance, and so is more visible as a contaminant.

[0044] With reference to section Figure 1 (d), this shows three rows (lines) of drinks. Each row comprises a triplet of 3 light beer drinks (labelled 1-3) and a triplet of 3 dark beer drinks (also labelled 1-3). Each triplet comprises 3 different types of glass - in particular the second glass in each triplet has a stem, while the third glass in each triplet has a reddish tint (so it has lower transmissivity for blue and green light).

[0045] In the top row of Figure 1 (d), the drinks are uncontaminated. In the middle row of Figure 1 (d), the drinks are contaminated with a paracetamol tablet. In the bottom row of Figure 1 (d), the drinks are contaminated with a multi-vitamin tablet. The addition of the multi-vitamin tablet generally results in a visible change in appearance of the drinks. For example, the first two glasses containing the dark beer with the multi-vitamin tablet have changed from translucent brown to a more opaque coffee colouring with a similar change also apparent for the third glass of dark beer. An analogous result is obtained for the light beers, in that the multi-vitamins produce a cloudy (low optical depth) and yellower appearance. In contrast, the glasses containing the paracetamol are very similar to the glasses without contamination, such that the paracetamol contamination is difficult to detect using just a routine visual inspection.

[0046] As described herein, a smart ring 120 may be used to detect when a drink has been spiked with (for example) small size drugs which are readily available, such as aspirin and paracetamol, or sedatives such as mirtazapine (Mirtazapin) (available on prescription). The Hedgehog system described herein may therefore be used to detect medical compounds such as pharmaceuticals at different concentration levels within a drink. Further, we describe below an investigation into drink spiking in various real-life situations and analyze different factors that can influence the drink spiking estimation, including: different cups, other soluble pill types, types of drinks, and ambient luminosity levels. The results from this investigation show that the Hedgehog sensing system operates in a robust manner across different conditions.

[0047] Figure 2 is a schematic diagram showing an example of a (Hedgehog) sensing system 200 as described herein to detect whether a drink has been spiked. In particular, the user is holding a glass 110 which contains a drink 112. The glass is at least partly transparent to allow the sensing system 200 to perform an optical analysis of the drink 112 to look for any spiking of the drink 112.

[0048] The sensing system 200 comprises a smart ring 120, a smart watch 210, and a smartphone 220. The smart watch is worn on the wrist, held in place by a suitable form of wristband, such as a Velcro strip. The smart ring 120 is shown worn on the fourth finger 122 of a user (i.e. the finger next to the little finger), however, the smart ring 120 may potentially be worn on any finger (including the thumb) according to user comfort, compatibility with other rings worn by the user, and so on. Various possible locations (fingers) for wearing the smart ring 120 are shown in Figure 4 (left-hand portion). As discussed in more detail later, the different fingers may give different levels of performance for the sensing system 200.

[0049] The smart watch 210 may act as an intermediary between the smart ring 120 and the smartphone 220, whereby bi-directional communications may be performed between the smart ring 120 and the smart watch 210 as indicated by arrow A1 , and these may be complemented by bi-directional communications between the smart watch 210 and the smartphone 220 as indicated by arrow A2. The smartphone 220 may further communicate with a remote server (not shown in Figure 2), such as by using the Internet and / or mobile data communications as supported by existing devices (smartphones and so on).

[0050] In some implementations, the smart ring 120 may also support bi-directional communications directly with the smartphone 220, as indicated by arrow A3, without passing through smart watch 210. In some cases, the sensing system may therefore comprise just the smart ring 120 and the smartphone 220 but without the smart watch 210. In these cases, the functionality of the smart watch 210 may for example be integrated (incorporated) into the smartphone 220, thereby allowing the smart watch 210 to be dispensed with. Alternatively, the smart watch 210 may be retained, and the sensing system 200 may implement both direct (arrow A3) and indirect (arrows A1 and A2) communications between the smart ring 120 and the smart phone 220, such as according to system settings, user preference, etc. Some implementations may omit the smartphone 220. In this case, the relevant functionality of the smartphone 220 may be incorporated into the smart watch 210. It will be appreciated that these different configurations are provided by way of example only, and are not intended to be limiting. The communications between the components of the sensing system 200, namely the smart ring 120, the smart watch 210 and the smartphone 220, are generally performed by wireless communications, such as Bluetooth, WiFi or near-field communications (NFC). Each of the components of the sensing system 200 may support multiple forms of wireless communication, and a particular form of communication for use at any given time may be selected given the properties of the other components in the sensing system 200, user preferences, and so on. Furthermore, the sensing system 200 may potentially use different forms of communications to or between different components. For example, communications between the smart ring 120 and smart watch 210 (as per arrow A1 ) might be performed with NFC and communications between the smart watch 210 and the smartphone 220 (as per arrow A2) might be performed using Bluetooth.

[0051] In terms of general operation, the glass 110 is held in the hand of the user in normal fashion, such that the smart ring 120 is in contact with the glass 110. This allows the smart ring 120 to emit light into the glass 110 for reflection and scattering by the drink 112 in the glass 110. (It is noted that reflection and scattering are two distinct physical processes, but for present purposes, we will use the term reflection to cover both reflection and scattering, unless the context clearly indicates to the contrary). Some of the reflected light is detected by the smart ring 120 and the measured properties of the reflected light are compared to the reference properties (fingerprints) mentioned above. If the measured properties deviate from the reference properties, this may indicate the drink 112 has been spiked.

[0052] In this latter situation, in which it is indicated that a drink has been spiked, the owner of the drink (and user of the Hedgehog sensing system 200) may be sent an alert message in one or more ways. For example, the smart ring 120 and / or the board or smart watch 210 might illuminate a warning light, issue an audible alert (e.g. beep), and / or perform some vibration. Accordingly, the alert may use a visual channel, an aural channel, and / or a haptic channel. Additionally or alternatively, other forms of notification or alert may include sending a message (email, WhatsApp, text) to the smartphone 220 or some other device of the user. The details of the notification (alert) may typically be set according to user preferences. As discussed above, communications associated with such notifications may be sent as appropriate from one device in the sensing system 200 such as the smart ring 120, smart watch 210 and / or smartphone 220 to another device in the sensing system (and / or to a remote server not shown in Figure 2).

[0053] Figure 3 is a schematic diagram showing an example of a smart ring 120 such as may be used in the Hedgehog sensing system 200 of Figure 2. The smart ring 120 is used to perform an optical analysis of the drink in the glass 110 and is fitted with a block 0 310 which comprises a number of active components and which extends azimuthally around a portion of the ring circumference. The block 0 310 of active components may be embedded in the ring. The remaining circumference of the ring 120 is indicated as comprising blocks 1 through to block N. Note that in some implementations, this remainder portion 320 may represent a single block (in effect, N=1 ). This remainder portion 320 may be passive in the sense that it provides structural support for block 0, but it does not include active components for supporting the primary functionality of the smart ring 120 (detecting a spiked drink).

[0054] For the smart ring 120, the user will generally (and without limitation) position the smart ring so that block 0 310 is directed inwards from the front back of the hand, i.e. pointing inwards from the palm of the hand, rather than out of the back of the hand. This positioning may be achieved by selecting an appropriate azimuthal angle for the ring at the time it is put onto the finger, and / or by adjusting the azimuthal angle of the ring after the ring has been put onto the finger. With this configuration of the hand, the block 310 will generally be directly against the exterior wall of the vessel 110 as the user holds the vessel to drink from, thereby supporting the transmission of light from the smart ring 120 into the vessel 110 (and vice versa). Note that in contrast to some approaches for detecting a spiked drink, the present approach does not involve any insertion of the ring, sensor, or some reagent, etc. into the drink 112 itself.

[0055] In some implementations, the block 0 310 of active components may be embedded into the ring 320. In other implementations, the block 0 310 may be mounted into a gyratory spin structure that moves freely in an azimuthal direction around the circumference of the ring 120. In this type of implementation, the remaining portion 320 may form a complete ring (360 degrees) and the block 0 310 may be configured to move around this ring so that the active components of block 0 310 are positioned at the appropriate azimuthal angle (such as to face out towards a drink held in the hand).

[0056] A further possibility is for at least some of the active components of block 0 310 to be repeated, for example at every 120 degrees around the smart ring 120. This may help a user position one of the blocks 0 310 at the appropriate azimuthal angle for use with less adjustment than might be involved if there is only a single block 0 310 of active components (although repeating the active components in such a manner would raise the production costs for the smart ring 120).

[0057] The active components shown in Figure 3 include two light emitting diodes (LEDs) 170 which act as light sources that emit light which is incident on the drink 112. In some implementations, at least one of the LEDs may produce monochromatic light of a single fixed colour, for example, only red, yellow or only white. Figure 3 shows a smart ring 120 as including red, green, blue (RGB) LEDs which allow the light source(s) to emit light of different colours (including white) according to the relative strength of the red, green and blue components. It will be appreciated that coloured (RGB) LEDs offer greater discriminatory power in terms of detecting whether a drink has been spiked.

[0058] As shown in Figure 3, a photoreceptor 160 (shown as a rectangle) is located between two opposing RGB LEDs 170, each represented by a circle. The photoreceptor 160 receives and captures light emitted by the RGB LEDs which has been reflected by the vessel 110 and the drink 112 therein, or by any contaminant (spike component) therein, back towards the active components of block 310, in particular towards the photoreceptor 160. The combination of the RGB LED(s) 170 as a light source and the photoreceptor 160 as a light receiver may be regarded as a light sensor (or light sensors, according to configuration). In operation, information concerning (i) the two light source signals produced by the two RGB LEDs 170 and (ii) the reflected light received by the photoreceptor 160 may be processed to reduce noise.

[0059] In some implementations, the photoreceptor 160 may have optical sensitivity across a relatively wide waveband, for example, encompassing the full range of wavelengths emitted by the light source 170. In some cases, the photoreceptor 160 may only measure the intensity of the reflected light without spectral (colour) information. In other cases, the photoreceptor may for example have RGB receptors, which can then be used to determine the relative amounts of red, green and blue in the received light. Having such increased information (colour) about the light incident on the photoreceptor 160 again may offer greater discriminatory power in terms of detecting whether a drink has been spiked.

[0060] The configuration of the light source (RGB LEDs 170) and light detector (photoreceptor 160) shown in Figure 3 is by way of example only and without limitation. Many further configurations will be apparent to the skilled person, including having fewer or more light sources, having more light detectors, having one or more light sources and / or light detectors that operate outside the visible range (such as in the near infrared range), having multiple light sources differing from one another (such as in wavelength), having multiple light detectors that differing one another (such as in wavelength), and so on.

[0061] In some implementations, the two (or more) RGB LEDs may be configured to produce light of the same colour as each other, or light of different colours - this latter configuration may provide better diagnostic information for determining whether or not a drink has been spiked. In some implementations the RGB LEDs may transmit light of a first colour and of a second colour different from the first colour. This may be achieved by having each RGB LED emit a different colour, for example, one RGB LED might emit blue and another RGB LED might emit red. Another possibility is that both RGB LEDs may emit a first (single) colour, e.g. blue, and then subsequently emit a second (single) colour, e.g. red. In some implementations, one or more additional colours (third, fourth, etc) may also be utilised.

[0062] The proportion of reflected light detected for the first colour relative to the proportion of reflected light detected for the second colour for a given drink can contribute to a reference characteristic (fingerprint) for that drink. If there is a subsequent change in the proportion of reflected light detected for the first colour relative to the proportion of reflected light detected for the second colour, this in effect indicates some change in the overall colouring of the drink in the container and might indicate contamination (spiking) of the drink. Accordingly, the use of LEDs (or other light sources) which can produce different colours provides additional discriminatory information for detecting spiking compared to purely monochromatic illumination.

[0063] Figure 3 further depicts ambient light 155, in other words background lighting generally provided at the location of the user of smart ring 120, as well as a luminosity sensor 180 used to measure the level of background lighting. In some cases the luminosity sensor 180 may detect the overall level (brightness) of background lighting; in other cases, the luminosity sensor 180 may also detect colour (waveband) information, such as by using an RGB detector (analogous to an RGB photoreceptor as discussed above). Typically, the luminosity sensor 180 is used to measure the intensity of the surrounding (ambient) light because this ambient (background) brightness level affects the measurements collected by the smart ring 120. Such measurement of ambient light may be performed at the same time as the measurement by the RGB LEDs 170 of the light reflected from the drink. This synchronisation of measurement between the photoreceptor 160 and the luminosity sensor 180 may be particularly significant when the venue has ambient lighting that typically changes with time (e.g. mood or disco lighting). Accordingly, the measurement of ambient lighting is useful for estimating the background noise levels associated with the measurements by the photoreceptor 160 and hence for determining whether a drink has been spiked.

[0064] In addition to the data obtained from the two or more RGB LEDs 170, the photoreceptor 160 and the luminosity sensor 180, which may all contribute to the analysis of the optical measurements to determine whether a drink has been spiked, the smart ring shown in Figure 3 further includes an accelerometer 165 and / or a pressure / proximity sensor 175. The pressure / proximity sensor 175 may comprise a pressure sensor, a proximity sensor, or both a pressure sensor and a proximity sensor (whether provided as separate devices or integrated together into the same device). The accelerometer sensor 165 can be used to compensate for motion, different angles and different distances to the target drink container. In a typical situation, the smart ring 120 may be held firmly against the drink vessel 110. The accelerometer 165 detects movement of the hand, and hence of the drink vessel which is held in the hand. The accelerometer may detect (for example) that the drink vessel 110 has been tilted, which may change the level of reflected light, for example because such tilting has moved the liquid 1 12 in the vessel 110 away from the sensor 160, 170. The movement of the vessel may also impact any particulate matter, for example, such particles may be agitated within the liquid rather than settling to the bottom (see the plot of Figure 7, section (b), for further consideration of this). With increased testing, the impact of such movement as detected by the accelerometer 165 on the drink sensing can be better understood and compensated for. For example, in a simple configuration, the smart ring 120 might only acquire a sample to test for spiking when the movement as detected by the accelerometer 165 is relatively small (below some threshold), thereby eliminating noise that might be caused by sudden movements of the vessel (such rapid movement may occur, for example, if the drink is being held by a person who is dancing).

[0065] A proximity sensor 175 may be used, for example, to confirm whether the sensor 160, 170 is close enough to the glass vessel 110. The control system within the smart ring may use this information to determine when to acquire samples of reflected light - i.e. such samples would only be acquired when the smart ring 120 and the glass vessel 110 are in close proximity to one another to provide reliable measurements. Amongst other things, this reduces the battery consumption (and so prolongs battery life) of the smart ring 120, in that the smart ring only emits light when the smart ring is detected by the proximity sensor 175 to be close enough to the glass vessel to obtain a useful and reliable measurement of the reflected light. In addition, the measured value of reflected light, such as determined by photoreceptor 160, is generally sensitive to the distance of the sensor 160, 170 from the drink container 110. The proximity sensor 175 can detect this distance which therefore allows the measured values to be compensated for this distance by calibration to a common, standardised distance (such as zero).

[0066] A pressure sensor 175 may detect when the light sensor 160, 170 is in contact with the drink; this may then prompt the light sensor to start emitting and detecting reflected light to perform a spiking test. In broad terms, having the light sensor firmly against the drink vessel 110 may help to reduce (cut out) extraneous noise, which may lead to more reliable results about the presence of any contaminants in the drink 112. In some cases, the sensing system may be calibrated according to the applied hand-grip strength which can also lead to more reliable results. In some cases, the smart ring 120 might also (or additionally) provide feedback to the user when such a level of pressure is applied - e.g. a light illuminating and / or the device making a ‘beep’ or other similar noise. Having such a feedback mechanism may be used to derive hand-grip strength as an object is squeeze by an individual. As shown in the right-hand portion of Figure 4, the smart ring may be configured to support a measurement of hand-grip strength by using an additional compressible component. Such a facility may already be available for some smart devices directed to the medical market. This measurement facility may then be used (re-purposed) by the smart ring 120 to assess the likely pressure to be applied by a given user of the sensor to the glass, and to calibrate the processing of subsequent measurements of reflected light accordingly.

[0067] The smart ring 120 also generally contains standard components, such a processing facility for executing program instructions to perform functionality as described herein, memory or storage for holding program instructions and data, a rechargeable battery and communications and recharging facilities (not shown in Figure 3). The operation of the communications facility has generally been discussed above in relation to Figure 2, see for example the arrows A1 and A2. The smart ring 120 may further include one or more user interface components (not shown in Figure 3), including output components such as an additional light and / or a speaker which may be used to indicate information such as a spiked drink, low power level, and so on. The user interface components may further include (for example) an on / off button, a button used for configuration, and so on.

[0068] Experimental Study

[0069] As described herein, the Hedgehog sensing system 200 provides an innovative pervasive sensing solution for identifying drink spiking. The Hedgehog system re-designs and re-purposes optical sensors on wearables, such as smart rings or smartwatches, so that they can be used to identify changes in light reflectance caused by small particles inside a drink. The Hedgehog system identifies drink spiking by first establishing a reference fingerprint of a drink and then monitoring for any deviation from this reference fingerprint. By developing rigorous benchmarks which take into account different types of drinks and pillsized doses of compounds that may be introduced into such drinks, it is demonstrated that the Hedgehog system can accurately and robustly identify drink spiking with an accuracy of up to 90% accuracy in detecting whether a drink has been tampered with. This approach therefore demonstrates a new use for wearables that can improve personal safety by helping to reduce the risk of consuming a spiked drink.

[0070] The experimental study was generally performed on an earlier version of the sensing system 200 than described above, for example with reference to Figures 3 and 4. In particular, the smart devices shown in Figures 3 and 4 include some additional or enhanced components compared to this earlier version of the sensing system, such as multiple coloured LEDs, an accelerometer, and so on. It will be appreciated that the smart devices shown in Figures 3 and 4 would therefore be able to match, and generally surpass, the performance obtained in this experimental study.

[0071] Identifying Spike Drinks

[0072] Figure 5 is a schematic diagram depicting an overview of the Hedgehog system 200 for assessing whether or not a drink has been spiked. Figure 5 includes the following sections: section (a) provides a high-level view of a smart ring 120; section (b) illustrates schematically the change in reflectivity resulting from contamination of a drink (spiking); section (c) indicates various locations where a user might wear the smart ring 120; section (d) illustrates the use of the smart ring to test (sample) the drink for contamination; and section (e) illustrates a sensing pipeline for analysing the data from the smart ring 120. In particular, section (a) shows two rings, the top one being a conventional smart ring. This top smart ring includes a sensor (including a light emitter) 134 which faces (and emits light) towards the centre of the ring. This type of smart ring is generally intended to investigate medical properties of the finger on which the ring is worn, and hence to discern certain health information of the person wearing the ring. By way of example, the sensor(s) may be used to collect measurements from an individual’s skin to derive metrics such as skin temperature or heart rate

[0020] . (For the avoidance of doubt, this type of ring, as shown in the top portion of section (a) of Figure 5, is not part of the Hedgehog sensing system 200 as described herein).

[0073] The lower portion of section (a) of Figure 5 shows a ring 120 which does form part of the Hedgehog sensing system 200 as described herein. The ring 120 includes a segment or block 310 including a light source such as the RGB LEDs 170 and a light receiver such as photoreceptor 160. Note that for this ring 120 in the lower portion of section (a), the active components of block 310 of the ring 120 are located on the outside (external surface) of the ring, facing radially away from the centre of the ring. As a result, the light sensor 160, 170 does not transmit light into a finger on which the smart ring 120 is being worn, but rather can transmit light onto / into an object which is outside (external to) the smart ring.

[0074] Accordingly, the light sensor(s) 170, 160 on the outer surface of the ring are able to collect light reflectance measurements from liquid containers. To make such a measurement, the light source(s) and light receptor(s) are generally placed against the surface of a liquid container holding the drink to be measured. The liquid container should be at least partly transparent or translucent to allow light from the light source(s) to enter and be reflected from the liquid 112. One important consideration is how to best position the sensor(s) 160, 170 to take obtain quality measurements. Various experiments are described herein regarding such positioning and these demonstrate that measurements can be robustly acquired using the smart ring 120.

[0075] The Hedgehog sensing system 200 exploits the principle of light scattering and reflectivity to assess changes in drinks [1 , 22]. When a user first acquires a drink, the user contacts the drink container with the smart ring 120 to measure the reflected light which in turn can be used to establish a reference fingerprint of the liquid 112 inside the glass. Thus Figure 5, section (b), shows two representations (left and right) of a drinking vessel 110 containing a liquid 112. In each case, the smart ring 120 contacts the drinking vessel 110 to perform the measurements. (The smart ring 120 is shown for clarity in Figure 5 as separated slightly from the drinking vessel 110 but in general the smart ring 120, especially sensor 160, 170 would be pressed against the surface of the vessel 110 when making a measurement). The left-hand representation of the liquid 112 is shown as clear and generally free from impurities. According, there is little or no scattering of light within the liquid itself, but rather reflection of the light at the surface(s) of the drink vessel 110 and refraction as light passes from one optical medium into another optical medium. These initial measurements of liquid 112 are used to create a set of fingerprints (reference parameters or base-line) which are indicative of no impurities.

[0076] The drink is continuously (repeatedly) checked using the smart ring 120 to monitor and confirm that the drink 112 remains un-spiked. For each check or repeated measurement, the measurements are compared against the reference parameters; if the measurements continue to match the reference parameters initially obtained for the drink, this is indicative of no spiking. In contrast, if the new measurements deviate from the reference parameters, this may be indicative of spiking.

[0077] The right-hand representation in Figure 5, section (b), depicts greater complexity in the optical paths within the liquid 112. In addition to the reflection and refraction described above, this greater complexity may be due, for example, to scattering by impurities which are now present in the vessel 110. This increased scattering may be measured by the sensor(s) 160, 170 as a change in the distribution of the light intensity values. If the measured light no longer conforms to the fingerprints (reference parameters) previously established, this is indicative of spiking. In practice, such monitoring for spiking may involve measurements (samples) collected over a sufficiently long period to allow the distribution to be estimated accurately.

[0078] Figure 5, section (c) shows a user hand with a smart ring 120. As indicated in Figure 5, section (c), the smart ring 120 may have a particular location (finger) according to user preference. Figure 5, section (d) shows the ring located on the fourth finger of the right hand. The active portion 310 of the smart ring 120 is generally located on the inside of the hand, so it is pressed against the surface of the vessel 110 when a user holds or lifts the vessel 110. This pressure of the smart ring 120 against the side of the drinking vessel can help to ensure more reliable optical measurements of the reflected light by sensor(s) 160, 170. In some implementations, multiple measurements are made while the glass 110 is in the left-hand representation of section 5(b) (unspiked). For example, 50-100 measurements may be acquired over a time period of 10-20 seconds in order to obtain a robust set of reference parameters or fingerprints. It will be appreciated that this number of measurements and corresponding sampling time is by way of example only and may be varied according to the particular circumstances. Once a suitable number of measurements has been made to accurately determine the fingerprints, the sensing system is then ready to detect future spiking of the drink 112.

[0079] The sensing pipeline used to process the optical signal received from the sensor(s) 160, 170 is illustrated in Figure 5, section (e). First, raw data of the light measurements is cleansed (processed), for example by using median filtering and convolution smoothing. Median filtering is typically used to remove outliers, while the convolution smoothing may remove or reduce some higher frequency components (typically noise). Next, a reference fingerprint is established from the initial set of samples and assigned to the specific liquid-filled container 112 of the user.

[0080] Successive test fingerprints are then obtained from subsequent optical measurements of the drink 112 and are compared to the reference fingerprint previously determined, to determine whether or not the drink has been spiked compared to the initial conditions for the reference fingerprint. This determination can be regarded as a classification problem which can be implemented, for example, using a machine learning classifier. Two classifiers which are relatively easy to implement have been tested, namely Random Forest (RF) and Gradient Boosting (GB).

[0081] The Beer-Lambert law expresses the linear relationship between the absorbance of light by a liquid substance and its concentration. It measures the amount of light that is absorbed when passing through the liquid at specific wavelength. Absorption A then is defined as A = log(IO / l), where I0 is the intensity of the incident light (intensity of the source laser diode) and I is the intensity of the transmitted light (value measured from the reflected light using the photo-resistor). As a drug compound increases, the concentration of the liquid changes, thus changing its overall absorption. Indeed, drug particles cause light to be scattered, modifying the amount of reflected light captured by the sensor. Absorption changes the values of reflected light collected by the sensor from different angles when analyzing the water in a transparent glass container (drink). Thus, absorption can provide a reference value on the type of liquid in the container, and the amount of drug that it was spiked with.

[0082] An enhanced analytical methodology is described below. The system employs light reflectance measurements from liquids to estimate whether a drink has been spiked. This process is carried out using a four-stage pipeline, which is outlined in Figure 11(a) to (d).

[0083] The first stage (a), sampling, involves collecting the light reflected from drinks using an optical sensing device. This device comprises a red-laser light emitter and a photo-receptor integrated into a smart ring, as described above. The light may be collected from the side of the container or from above.

[0084] The second stage (b), pre-processing, removes the noise from the collected measurements for example by using convolution filtering or other suitable de-noising algorithms. This pre-processing compensates for noise introduced by hand motion and variations in ambient light conditions. At least one condition sensor is provided to detect the condition of the sensing device, such as its motion or orientation, for example using an accelerometer, gyroscope and / or IMU (inertial measurement unit). Further condition sensors provided to detect the background luminosity in the environment. In everyday settings, the light values are affected by hand motion and variations in the ambient light, as shown in Figure 12(a) and 12(b), where light intensity profiles are shown for different regions of a social space, such as a bar or nightclub, where luminosity varies in different areas; AO is the bar area, A1 is the dance floor, A2 is the sitting area and A4 is across the 3 areas. Unless taken into account, these factors can hamper spiking detection. A calibration approach is included that adjusts the light values based on the accelerometer sensor and the ambient light intensity (Lux).

[0085] Figure 13 shows the processing pipeline for motion and Lux calibration. Light values are calibrated for motion in three steps. Before the first step, each motion axis and the light values are normalised. Normalisation can be done by for example by z -score, a dimensionless quantity that is used to indicate the signed, fractional, number of standard deviations by which an event is above the mean value being measured. This is to achieve a uniform calibration in case some light values cannot be represented.

[0086] Step 1 : Motion Projection. Light reflectivity is greatly affected by motion. Light measurements are impacted by user behaviour in different contexts, therefore the motion behaviour in different areas is checked before calibration. For example, in a nightclub setting, a person sways more frequently in the dance area than in the seating area, which leads to larger variations in light values in the dance area. As users behave differently in different areas, the motion distribution is analysed for each area.

[0087] An Empirical Cumulative Distribution Function (ECDF) is used to calibrate light values in each area of the nightclub based on motion data detected by the accelerometer. Since accelerometer data is not independently and identically distributed, ECDF representation can effectively quantify its distribution.

[0088] A ECDF is applied on each motion axis to find the cumulative probability distribution (Pmx,Pmy,Pmz) in the first step. Light values I are projected in each-axis ECDF to check the individual effect on light reflectivity. A new cumulative probability distribution vector is thus obtained (Pm (Z) = [Pmx (Z), Pmy (Z), Pmy (Z)]). These new probabilities indicate the effect of each motion axis on light-value distributions.

[0089] Step 2: Adjusted Weighted Averaging. The effect of each axis is combined in the second step. Classical machine learning models Random Forest (RF) and Gradient Boosting (GB) are used to explore the effect of each motion axis on the light values and obtain the average weights (l / l / x,l / l / y,l / l / z). A simple weighted average cannot be used to calibrate the light value cumulative distribution due to the heterogeneity of each motion-axis effect. The effect of motion-x is minimised by subtraction in the adjusted weighted averaging to get the calibrated cumulative probability representation Pm (Z) for light values.

[0090] Step 3: Inverse projection. The calibrated probability Pm (Z) is put into the inverse ECDF original light values to get the calibrated light values Zi. Since light values are normalised, for example by z-score, the calibrated value Zc can be recovered from Zi by original mean values and standard deviation. After motion calibration is performed, Lux calibration is done to obtain value Zc, where the various ambient Lux effects are factored in. Ambient Lux effects produce low-frequency noise with a high amplitude, thus the frequency domain is used to reduce the ambient Lux noise. 100 main frequency components were extracted according to Lux amplitude size. The choice of the number of dominant frequencies depends on the sample size. More frequency filtering can lead to a better result, but it may also cause overfitting. A digital notch filter is used to remove these specific components in the light-reflectivity frequency domain, and also filtering to remove phase distortion. An inverse Fast Fourier Transform is performed on the filtered light signal to convert it back to the time domain.

[0091] The third stage (c), feature extraction, aims to reduce the dimensionality of light values, such that efficient representations of what is a drink can be learned. Current implementation uses a deep learning autoencoders which are used to extract the latent representations from the pre-processed measurements. Other mechanisms such as sparse coding or self-organizing may provide similar functionality. An autoencoder is a type of neural network architecture designed to efficiently compress (encode) input data down to its essential features, then reconstruct (decode) the original input from this compressed representation. Autoencoders are trained to discover latent variables of the input data: hidden or random variables that, despite not being directly observable, fundamentally inform the way data is distributed. Collectively, the latent variables of a given set of input data are referred to as latent space. During training, the autoencoder learns which latent variables can be used to most accurately reconstruct the original data: this latent space representation thus represents only the most essential information contained within the original input. Most types of autoencoders are used for artificial intelligence tasks related to feature extraction, like data compression, image de-noising, anomaly detection etc. The feature extraction stage results in a lower-dimensional space that captures the representative features of spike and non-spiked drinks. Feature vectors are extracted from the data in a time window, T, by employing the encoder part of the autoencoder, which reduces the dimensionality of the data and captures the essential features. Mix-max normalization is applied to ensure that these values are comparable across different drinks and drugs in a consistent scale. A one-second sliding window with a 50% overlap is used to measure the variance of consecutive measurements. Different time window sizes may be used, for example with an increment of 1 second from 1 to 5 seconds.

[0092] Finally, the modelling stage (d) involves convolutional neural networks (CNNs), which are used to evaluate the how spiked a drink is, based on the latent representations from the autoencoders. These latent representations are then used to train a Convolutional Neural Network for binary (i.e. spike, non-spike) and multi-class (i.e. drug types) recognition. CNN models are trained using a leave-one-drink-out approach, whereby the models are trained and validated on all but one drink, with the left-out drink serving as a test set. Training and validation are conducted using 5-fold stratified cross-validation. Model evaluation employs standard classification metrics, such as precision and recall.

[0093] This approach requires robust and accurate feature extraction from reflected light measurements to reliably detect drink spiking. This is particularly challenging due to the variability in the types of drinks and potential drugs that can be utilized. The present system addresses these challenges using autoencoders, which extract representative features of signals, the latent spaces. Autoencoders are suitable here for use in data de-noising by determining whether the deviations of the latent representation of a signal from the learnt latent space of normal signals surpasses a threshold. The present system uses anomaly detection to detect drink spiking. This is achieved by comparing new observations against known models or patterns of normal drinks, i.e. drinks that are not spiked. This is established by applying statistical functions or pattern matching algorithms. The present system extracts the latent space from the pre-processed data, targeting the time window, T, where drink spiking is suspected. The latent space corresponds to the feature space representing significant deviations in the variance.

[0094] The light may be collected from the side of the container or from above. The motion detector may be an accelerometer or gyroscope to determine the orientation of the sensing device. In this way the processor can determine whether signals are being received from the side of the container or directly from the surface of the drink from above. The data can be processed to account for the signal coming through the container or directly from the surface of the drink.

[0095] Experiments

[0096] The Hedgehog sensing system 200 has been evaluated through four experiments which focus on: (i) characterizing different drink contents, (ii) detecting mixtures of soluble compounds, (Hi) detecting compounds and (iv) demonstrating practicability in the field. In all experiments, measurements are taken using a (test) smart ring that integrates at least a red light sensor. Measurements were taken from three different finger positions: little (fifth) finger (Position-1 ), middle (third) finger (Position-2), and index (second) finger (Position-3), as shown in Figure 4 (left-hand portion). In all experiments, 225 milliliters of the given liquid for testing were poured into a transparent glass (cup1 ), and we conducted 4 trials per drink per position, with each trial consisting of 6 one-minute measurement periods. The data were collected for one minute to ensure the data is representative and to mitigate the impact of micro-hand movements, re-adjusting of the grip, and other sources of noise. Further evaluations were performed to consider the effect of ambient luminosity, the use of different cups and the use of different drink types. More detailed information about these experiments is provided below.

[0097] A smart ring prototype that embeds the light sensors on a ring was constructed. This smart ring prototype was connected to a computing board that was used to analyse the measured light values; see Figure 1 , section (a). As noted above, this prototype uses a wireless M5StickC PLUS ESP32 development board which could (inter alia) control the sampling frequency of the light sensor 160, 170 and upload the samples to a server in real time. Another implementation might for example use a smartphone 220, tablet, or similar for analysing the measurements (rather than using a server).

[0098] The M5StickC Plus contains an inbuilt Wi-Fi connection facility, battery supplies (120 mAh @ 3.7V), and an LCD screen to externalize the activities of the board. It is lightweight (21g) and portable, such that it can be placed in a wristband (65 mm *25 mm *15 mm). A light sensor was integrated into the system through a separate wire that is attached to a plastic ring. The plastic ring is made from an elastic wire which makes it easy to adjust the ring to different finger sizes.

[0099] One implementation of the prototype used an optical sensor comprising a red laser diode (650 nm, 5 mW, 3 - 5V) and a photo-resistor (5MQ) to obtain the optical measurements. The photoresistor measures the intensity of light reflected back from the beam produced by the laser diode. The photoresistor captures analog voltage measurements which are converted to digital voltage representations. The ADC output value (analog to digital conversion with 12-bit resolution) is used as the physical unit for reflected light intensity. The sample rate is configured at 5 Hz frequency, and on average, 50 samples were used to characterize a drink with light measurements at 97.5% confidence.

[0100] (i) Characterizing Drinks: In an initial stage it was assessed whether light reflectance measurements can be used to characterize and identify different types of drinks. We considered six commonly available drinks that can be easily manipulated in social contexts

[0017] . The drinks are shown in section (b) of Figure 1 and comprise (A) water, (B) carbonated soft drink (cola), (C) beer (5.2% alcohol), (D) milk, (E) cold tea, and (F) grape juice. (We refer to this setup herein as DRINKS).

[0101] (ii) Soluble Compound: The second experiment (MIXTURE) incrementally mixes a soluble compound - instant cocoa powder - into milk. The powder was added in increments of 2 g until a total of 10 g had been added (see Figure 1A), leading to an increasing suggestion of brown from the 2g glass to the 10g glass. After each increment, the mixture was stirred for about a minute and left untouched until the liquid had reached a stable state (no motion) for light fingerprint measurements.

[0102] (iii) Medical Compounds: The third experiment (PILLS) evaluated water that has been spiked with medical pills, see Figure 1 , section (c). Three different pills: aspirin (500 mg), paracetamol (500 mg) and mirtazapine (30 mg) were evaluation. The first two are generic, freely available, over the counter painkillers. The last one is an antidepressant that has mild sedative effects and can only be obtained with a prescription. The pills were ground into powder and mixed with a glass of water. The resulting mixtures were stirred thoroughly prior to taking measurements.

[0103] (iv) Practicability: The final experiment was performed to assess factors affecting the light measurements. These experiments only considered measurements from the middle finger (position-2) as the other experiments demonstrated this position provided the best performance (see below). The experiments were performed with different cups, pill types, types of drinks, and ambient luminosity levels. Other properties or situations investigated were: transparent and translucent glass cups: cup1 (2 mm thickness, baseline), cup2 (1 mm thickness, goblet) and cup3 (2 mm thickness, tinted cup, reddish-brown); additional soluble pills: paracetamol (500 mg, no color change) and multivitamin (color change); ambient luminosity: indoor dark (Dark), indoor ambient light (IAL) and outdoor ambient light (OAL); and drink types: light beer (4.7% alcohol) and dark beer (5% alcohol); see Figure 1 , sections (c) and (d).

[0104] Results

[0105] Drink characterization: demonstrating that individual drinks have light reflectance fingerprints that are unique to their contents is important for showing that reference fingerprints can therefore be used to determine whether or not the contents of a drink have been tampered with.

[0106] Figure 6, sections (a)-(c), shows the sensed light value (y-axis) for locating the ring on different fingers for the specified drinks (x-axis), namely position 1 for section (a) (fifth finger), position 2 for section (b) (third finger) and position 3 for section (c) (second finger). There are clear differences between the drinks, but these also differ according to the ring position. Kruskal-Wallis tests

[0010] demonstrated that light reflectivity in different positions could characterize different drinks that are poured into a standard transparent glass as follows: Position-1 : %2= 131.12, q2= 0.94, p < .05; Position-2: y2= 131.12, q2= 0.94, p < .05; Position-3: x2 =131 .93, r = 0.92, p < .05. Pairwise post-hoc comparisons for each sensor position and drink indicated significant differences in all but three cases - the exceptions were cold tea and grape juice measured on the little finger (%2= -1.30, p > .05), and cold tea and water measured on (i) the middle finger (%2= 0.35, p > .05) or (ii) the index finger (%2= -0.16, p >.05). The index (second) finger and the middle (third) finger typically result in a stronger grip than the little finger, which may improve the quality of the measurements by having the sensor pressed more strongly against the glass or other drinking vessel. Even for the index and middle finger positions, the light value readings may depend on how the container (vessel) is held and how exactly the fingers are positioned. The Hedgehog system 200 is able to incorporate measurements (samples) from multiple different contact points and angles over time, and taking advantage of this helps to improve the robustness of the measurements. The results also demonstrate that while differences in contents can be clearly observed, identifying the exact liquid type may be challenging. Nevertheless, as demonstrated below, the granularity of these differences is sufficient for use in identifying drink spiking.

[0107] Aggregated drink mixture: Figure 6 sections (d)-(f) demonstrate how the fingerprint of milk changes as a dissoluble compound (cocoa powder) is added. The y-axis again represents a sensed light value (but with different scaling), while this time the x-axis represents the amount of added cocoa powder. The sections (d), (e) and (f) correspond to ring position 1 , 2 and 3 (matching sections (a), (b) and (c) respectively). A Friedman test [6] verifies the differences in reflectance are statistically significant for all fingers (Position-1 : v2= 115.26, W= 0.96, p < .05; Position-2: %2= 111.71, W= 0.93, p < .05; Position-3: x2= 120, W = 1, p < .05). Pairwise post hoc comparisons using Wilcox-Bonferroni tests

[0018] further showed that all differences in light values are statistically significant for all compound amounts (p < .05) and each finger. These results were also supported by visual inspection as the drink became darker, and demonstrate that the Hedgehog system is capable of detecting changes in drink characteristics (such as from spiking).

[0108] Drink spiking: Figure 7 shows more experimental data in sections (a) to (e). Sections (a) and (b) show results for four different contents, namely pure water, water + aspirin, water + paracetamol and water + Mirtazapine. For each of these contents, three reflectance values are presented, namely for position 1 , position 2 and position 3, shown as purple, green and blue respectively. For each of the four different contents, the results for positions 1 , 2 and 3 are ordered accordingly along the x-axis; also, for each of the four different contents, the light value of position 1 was highest in value, then position 2, and position 3 was lowest in light value.

[0109] For the light reflectance measurements of Figure 7, section (a), in which medical pills are added to water, both the 500 mg compounds (painkillers) and the 30 mg compound (antidepressant) are easily identifiable. Note that these pills are not designed to be soluble and on visual inspection it is possible to see small particles in the water. These particles cause the light to scatter in different directions, changing the way in which the light is captured by the photoreceptor and causing significant alternations in the overall light value (fingerprint) of a drink. From Figure 7 section (a), it can be seen that while the light intensity values vary across different finger positions for the smart ring, the changes in contents are consistently visible across all positions.

[0110] With respect to the position of the finger (1 , 2 or 3), the particles generally moved slowly and dissolved from top to bottom. Typically, each position saw over time the same the spiking behaviour from different views, subject to additional factors such as the way of holding a glass, the distance from the liquid surface to the holding position, how strong the glass is held, and how tight the ring is against the glass surface, and so on. This result is reinforced by experiments (see below) in which the compounds were left to sink for an hour, and then measurements were retaken. In this case, the drink characterization fingerprints returned quite close to their original value.

[0111] Friedman tests in all the sensor positions demonstrated significant differences for all the spiked drinks (p < .05). Placing the sensor on the middle finger (position-2) provided the best results overall (%2= 51 .05, W = 0.71, p < .05). Pairwise post-hoc comparisons (Wilcox-Bonferroni) showed that the differences in light values are statistically significant for all the pairs (p < .05) except for the two painkillers, aspirin and paracetamol in position 1 (Z = -0.04, p > .05) and position 3 (Z = -1.67, p > .05). These results indicate that drink spiking can be detected robustly but the exact compound may be difficult to identify, especially if the compounds have a similar structure, for example it may be difficult to distinguish between the two painkillers. The middle finger (position 2) represents the best individual position for detecting changes, but overall using measurements from multiple contact points is preferred as it improves detection when the spatial distribution of the particles changes. For example, the compound may initially be mixed with the drink but then slowly fall to the bottom of the glass as it interacts with the liquid. This suggests that even partial changes, such as the formation of sediments, may be detected as long as multiple contact points are used for the sensors.

[0112] Spiking over time: After a drink has been spiked, many compounds gradually sink to the bottom of the container over time. However, this does not hold for all substances as some can remain suspended, submerged or float on the surface of the liquid, depending on their density relative to the liquid and their chemical properties. We now show that gradual changes may also alter the light fingerprint of a drink sufficiently to allow the Hedgehog sensing system 200 to detect this change in contents.

[0113] With reference to Figure 7, section (b), this shows measured light values for the same four contents as Figure 7, section (a), namely pure water, water + aspirin, water + paracetamol, and water + Mirtazapine. For the pure water, the normal (unspiked) light value is shown. For each of the three contents that include water plus an added drug, three measurements are depicted in Figure 7, section (b). The first measurement, denoted S, is acquired directly after a drink has been spiked (S). The drink is then left motionless for an hour before further measurements are obtained, denoted (SN). The spiked drinks are then shaken before further measurements are obtained, denoted (SS), to investigate whether the light properties return to their original values from directly after spiking.

[0114] Figure 7 section (b) shows these results, including a (pure) water reference baseline as mentioned above to facilitate the comparison. After resting for an hour, the light fingerprints of the spiked drinks change, returning to a measured light value SN which is closer to the water baseline (compared with the original spiked value S). However, after the drink is re-shaken, the measured light value SS generally increases again, in other words, back towards the original spiked range S. This suggests that detecting spiked drinks which are left unattended for a considerable period of time should involve measurements from multiple contact points, or else such drinks should be shaken in order to provide a reliable indication of whether or not they have been spiked.

[0115] Classification performance: The Hedgehog sensing system 200 also supports binary classification of spiked drinks. A 5-fold classification was used to train two simple classifier models: Random Forest (RF) and Gradient Boosting (GB) (see https: / / en.wikipedia.org / wiki / Random_forest and https: / / en.wikipedia.org / wiki / Gradient_boosting respectively).

[0116] When the light reflectivity values are considered from only one finger position of the smart ring 120 to identify drink spiking, the average estimation accuracy is 89.71% from the middle finger (position 2). The index finger (position 3) results in a performance of 82.24%, whereas the little finger (position 1 ) results in lower accuracy of 48.64%. The reason for the poorer performance of the little finger is attributed to the grip of a user. Specifically, the grip force exerted by the little finger is typically smaller when a user holds an object such as a glass container 110; this results in the ring having weaker contact with the glass container than when the middle or index finger is used. When the light source (such as diode 170) is not pressed tightly against the container, this may lead to a gap between the diode and the outer surface of the container, which can then reflect some of the light back directly, i.e. without this light ever passing through the liquid 112 in the container 110. This directly reflected light represents a noise source for the light measurements. In contrast, for the index and middle fingers, which generally do not allow such a gap between the diode and the outer surface of the container, this noise is not present. The (smaller) difference in performance between the middle and index fingers may result from the middle finger again being more firmly planted on the surface of the container compared to the index finger, resulting in less noisy light measurements. These variations in performance may also be detected by examining characteristics of the light signals. The classification can target situations where the signal is clean, suggesting that over 80% accuracy is very much achievable provided poor quality signals are filtered out.

[0117] Practicability results: Additional factors that might influence the performance of the Hedgehog sensing system 200 have been investigated. Figure 7 section (c) shows how different forms of ambient light intensity impact the measured light values. Both indoor ambient light (IAL) and Dark appear to produce only a small variation in the measured light values. However, outdoor ambient light (OAL) may result in additional reflections causing higher measured light values that might impact the detection of spiking.

[0118] Figure 8 contains sections (a)-(f), each representing a separate plot. For all the plots, the y-axis corresponds to the measured light values. There are two rows of plots, the top row comprising sections (a)- (c) having light beer as the baseline drink, and the bottom row comprising sections (d)-(f) having dark beer as the baseline drink. Plots (a) and (d) were acquired using indoor ambient light (IAL); plots (b) and (e) were acquired in dark conditions (D); and plots (c) and (f) were acquired using outdoor ambient lighting (OAL). Each plot is divided into three portions, the left portion representing the baseline (light or dark beer), the middle portion representing the baseline with added paracetamol, and the right portion representing the baseline with added multivitamins. Each portion of each plot comprises three measurements for cup 1 (purple), cup 2 (green) and cup 3 (blue). Within each portion, the cups are ordered along the x-axis with cup 1 first, then cup 2, then cup 3.

[0119] Figure 8 therefore shows the results for different cups and drink types under different light conditions. The results largely mirror those described above, showing that spiked drinks can be distinguished from those that are not spiked across all conditions. The classification models were also retrained to analyze how the following features affect classification performance: the light intensity captured by the photoreceptor (L), ambient light intensity (U), drink type (B) and glass cup type (C). The model was trained and tested using both 5-fold cross-validation and leave-one-trial-out cross validation to assess the impact of the training data and the robustness of the models.

[0120] The results from this work are show in Figure 9 (Table 1 ) which details drink spiking classification performance (%). The features used for determining the classification are: light reflectivity value (L), glass cup type (C), beer type (B) and light intensity LUX (U). L is the light intensity value captured by the photoresistor. This value is tied to the configuration of the sensor, and can be normalized into different scales, such that comparable trends can be found. LUX is a standard unit to measure illuminance. In this case, LUX is used to monitor the luminosity of the environment where the experiments where conducted, such that it is possible to replicate the experimental testbed and results.

[0121] The glass type (C) is shown in section (d) of Figure 1 , namely 1 , 2 or 3, and section (d) also shows the two beer types (light beer and dark beer). Figure 9 gives results separately for different lighting conditions, namely IAL (indoor ambient lighting), dark (D), and OAL (outdoor ambient lighting). Figure 9 further gives results based on all the above lighting conditions (ALL), i.e. without splitting the results by lighting conditions. The results for each lighting condition, IAL, D, OAL and ALL comprise two lines, the first line showing the results obtained using features L, C, B (as defined above), and the second line showing the results obtained with an additional feature U (i.e. L, C, B and U together.

[0122] It is observed from Table 1 that variations between lit and dark environments are relatively small, indicating that the Hedgehog sensing system 200 is resilient to changes in the luminosity of the ambient environment. In particular, Table 1 shows that the overall average classification results using different ambient lighting are as follows: IAL: 92.25%; Dark: 99.30%; and OAL: 94.91%. The best performance is obtained when all ambient conditions are considered, resulting in an accuracy of ALL: 94.29%. In contrast, a model that considers only the light reflectivity values in a particular setting (but not the glass type C or beer type B) results in about 61.65% accuracy; such a model needs additional information, e.g. using the information specified in Table 1. As more information (variables) are added, e.g., room luminosity and so on, the classification accuracy becomes better as shown in Table 1 , which shows some of the most important combinations for producing the highest accuracies for the ring to be used in practice.

[0123] Green Light Re-Purposing

[0124] Another implementation of the Hedgehog sensing system 200 has been developed by re-purposing a commercial off-the-shelf smartwatch to capture light reflectance. This re-purposing uses a Samsung Gear S3 Frontier smartwatch that integrates two green LED lights and a photo-receptor which are used for measuring heart rate data. This re-purposing involves wearing the watch in effect inside out, so that the clockface (time) was directed inwardly to the wrist, while the green light sensor was then directed radially out from the wrist, so that it could be placed against the surface of a glass 110 or other suitable vessel.

[0125] The re-purposed smart watch was used to repeat the experiment in which different pills were mixed with water. In particular, this investigation mixed aspirin (500 mg) and paracetamol (500 mg) into water to generally replicate the experiments shown in Figure 7, sections (a) and (b) (but without replicating the addition of Mirtazapine). Measurements of reflective light value were acquired from both the side of the glass container 110 and also from the bottom (underside) of the glass container.

[0126] Results: Figure 7 sections (d)-(e) shows the results for the smartwatch experiments. As for Figure 7, sections (a)-(b), the y-axis shows the measured light values. There are three data points indicated in each of sections (d) and (e), the left relating to (pure) water, the middle relating to water and aspirin, and the right relating to water and paracetamol. The measurements of Figure 7, section (d), were obtained by illuminating the liquid from the side of the drinking vessel 110, while the measurements of section (e) were obtained by illuminating the liquid from the bottom (underside) of the drinking vessel. It can be seen that for Figure 7, section (d), adding the contaminants (aspirin or paracetamol) reduces the detected light output from the glass (compared to water), while for Figure 7, section (e), adding the contaminants (aspirin or paracetamol) increases the detected light output from the glass (compared to water). This is related to the compounds (contaminants) moving from top to bottom after they are introduced into the top of the glass and the location of the sensor. In the case of Figure 7, section (d), the sensor is attached to the side of the glass, whereby the light values are largely dependent on the changes in the drink’s properties, such as colour, visibility, density, etc. However, in the case of Figure 7, section (e), the sensor is at the bottom of the glass. The glass has greater thickness at the bottom resulting in higher reflected light values. Furthermore, if the drug dissolves slowly, the sediment at the bottom of the glass further contributes to the reflected / scattered light output when the data is collected at the bottom as for Figure 7, section (e).

[0127] The results in Figure 7, sections (d) and (e), broadly mirror those obtained for red light with the smart ring prototype. This indicates that off-the-shelf sensors can be readily re-purposed to detect spiked drinks. The Friedman test indicated significant differences for measurements collected at the side of the glass: (%2= 4796, W = 1.0, p < .05) and at the bottom of the glass (%2= 5598, W = 1.0, p < .05). In both cases, pairwise post-hoc comparisons using the Wilcox-test (with Bonferroni correction) indicated significant differences for all pairs (p < .05) in both conditions. This implies that the position of the sensing device relative to the container may affect measurement. One way of addressing this is to incorporate inertial sensors (such as may be provided by accelerometer 165) into a smart ring 120 or smart watch 210 since the inertial sensors can be used to determine the orientation of the sensing device.

[0128] Discussion

[0129] As described herein, light reflectivity may be used both to characterize drinks poured into different transparent glasses and also to accurately detect when a drink has been spiked. Such a sensing device may accommodate frequent changes in the ambient environment. For example, in a social gathering an individual may move through rooms that are differently lit and be exposed to different types of drinks and containers. To ensure accurate performance for detecting spiked drinks in such a dynamic setting, the derived fingerprints should be reliable to minimize unwanted false positives, especially having regard to diverse real-world settings such as (for example): diverse luminosity, different hand movements and different sensor angles. A device may also be developed to perform the measurement process from the top of the glass (given that not all glasses are transparent [4], e.g., mugs).

[0130] Other methods: The Hedgehog sensing system 200 utilizes optical measurements taken during natural hand interactions with drink containers using a smart ring worn by a consumer. Existing sensor technologies such as RF sensing

[0014] and / or audio sensing

[0012] may also be re-used (re-purposed) to help detect if a drink has been spiked. In this respect, a key benefit of the Hedgehog sensing system 200 is it can be used “out of the box”, i.e. without the need for (specialized) measurement setup, system configuration, and so on. Furthermore, some wearable devices (e.g. smart ring, smart watch) already include sensors which could be used for monitoring drink spiking as described herein. From a commercial perspective, this greatly facilitates the adoption and scale-up of a Hedgehog sensing system 200.

[0131] Chemical effects of spiking: The presence and quantity of small particles in air are commonly detected using light scattering, e.g., this approach may be used, for example, to detect particulate matter in air and to identify plastic pollutants

[0022] . There is interest in developing models to identify chemical effects due to tampering, e.g., using light scattering to detect particles that gradually sink to the bottom or the formation of crystallized structures or to detect soluble compounds that are mixed with drinks by stirring them.

[0132] Usage performance: Drink spiking can occur in a matter of seconds and individuals only have 5 to 20 minutes before the first symptoms appear. This suggests that the Hedgehog sensing system should provide a warning or alert of a possible spiked drink as quickly as possible. The measurement time of the approach described herein can be reduced by increasing the sampling frequency of the sensors or by integrating models that predict the behavior of compounds in drinks - these then provide a quicker result to a user. Other alternative methods to improve the safety of individuals may be used to connect the Hedgehog system 200 with third-parties, e.g., police or relatives. For example, the Hedgehog sensing system 200 may automate the process of requesting third-party help immediately when it determines that a user has ingested (or otherwise contacted) a drink which has been tampered with.

[0133] Complex drink mixtures and other factors: Given the large spectrum of different drinks, drug substances and combinations, it can be challenging to determine whether an unknown drink has been spiked or not. Further work may investigate drink types, different drug dosages, different alcohol percentages and so on. In addition, the protection of drinks that combine different sources, e.g., cocktails, may involve an initial characterization of their contents before drink spiking events can be detected. Moreover, other factors may influence the detection of drink spiking. For example, objects inside the drinks such as ice, carbonation level, spoons and / or straws may significantly modify the optical properties of the drinks; again, an initial characterization might be adopted. A further area of interest is to apply this approach to hot beverages such as coffee, since hot contents may be better at dissolving foreign compounds.

[0134] Databases and models: Similar to manufacturer databases and online dictionaries that provide detailed specification about drugs, the Hedgehog sensing system 200 provides a way to collect data about drink spiking. This data may then be accumulated into a drink spiking databases (datasets). Such a database may provide details about the type of drink and the substance(s) that were used, such as drug type and concentration. These databases may also collect information from symptoms faced by an individual who may have consumed a spiked drink. In parallel to this, Al models may be trained from this data in order to provide robust detection and actionable recommendations. As wearable devices continue to evolve, easy access to such databases and models may help reduce false positives for spiked drink for new wearables.

[0135] Further information

[0136] As described herein, a sensing system and method are provided that can be integrated into wearables for identifying drink spiking. It has been experimentally demonstrated that such a sensing system and method can be used to identify different types of drinks from specific light reflectivity fingerprints, and to identify drink spiking robustly and accurately with approaching 90% accuracy. The disclosed system and method may be readily implemented using existing off-the-shelf commercial smartwatches or other similar devices, and work with different cup types, drinks, ambient luminosity conditions, and so on. Accordingly, the system and method disclosed herein harness wearables to support personal safety and identify spiked drinks.

[0137] As described herein a wearable sensing device may be used for determining whether or not a drink has been spiked. The sensing device may include a sensor for holding against a container including the drink. The sensor includes one or more light sources to be placed against the side of the container to transmit light at first and second wavelengths into the container, wherein the first wavelength is different from the second wavelength. The sensor further includes at least one photoreceptor for placing against the side of the container to receive light which has been transmitted into the container and reflected (scattered etc) out of the container. The intensity of the light received by the photoreceptor for the first and second wavelengths may be compared with reference characteristics for determining whether or not the drink has been spiked.

[0138] The wearable sensing device may be provided by a smart ring with the light source(s) configured to transmit light in a direction radially outwards from the smart ring through the side wall of the drink container. For example, the smart ring may be worn with the sensor located towards the front (palm) of the hand for the sensor to make contact with the drink container when a user takes hold of the container. The smart ring may be further provided with components such as a processor, memory (storage), communications, a user interface and so on, as for existing smart rings, to provide the platform for the operations described herein.

[0139] Some implementations include first and second light sources each comprising a red, green, blue light emitting diode (RGB LED). The first and second light sources may be located on opposing sides of the photoreceptor to provide a compact and effective configuration. Having multiple colours available can help in the determination of whether or not a drink has been spiked, such as by detecting a change in (i) the ratio between the transmitted light at the first and second wavelengths compared with (ii) the ratio between the received light at the first and second wavelengths. The use of different coloured light sources potentially allows the sensor to discriminate between (i) a general rise in light absorption by the drink (so a lower intensity of reflective light) and (ii) a change in colour of the drink.

[0140] Some implementations may include a proximity sensor and / or a pressure sensor to determine contact (or potentially near contact for a proximity sensor) between the sensor and the container. Such (near) contact may be used to initiate transmission of light from the light source(s), which can increase battery life by preventing light transmission when the sensor is too far from the drink container to be reliable. Such (near) contact may also be used to assist in providing a good interface between the sensor and the drink container to reduce or prevent extraneous light introducing noise that may interfere with the measurement of the received light to determine if a drink has been spiked. The sensor may further include an accelerometer to determine the orientation in space of the wearable sensing device; this information may also assist in detecting whether or not a drink has been spiked. The sensor may further measure ambient light levels to compensate where appropriate for any change in such ambient light levels when determining if a drink has been spiked.

[0141] The device may include a processing facility to analyse the intensity of the light received by the photoreceptor for comparison with the reference characteristics to determine whether or not the drink has been spiked. The device may further include a communications facility (typically wireless) to transmit data on the intensity of the received light received. This data may be compared with the reference characteristics to determine whether or not the drink has been spiked.

[0142] The reference characteristics may be determined from the specific drink to be consumed by the user at the initial time of purchase. In other cases, the reference characteristics may be predetermined, for example, by a particular venue or chain of venues for some or all of the drinks that they sell. In this latter case, the reference characteristics may be downloaded to the wearable device by customers of the venue, thereby avoiding an initial measurement phase for each drink. A machine learning model may be used to determine (train) the reference characteristics of the drink, whether such determination is performed by a user or by a venue. In the latter case, the training may include drinks with known contaminants to allow the model to develop a better classification between uncontaminated and contaminated drinks.

[0143] ★★★

[0144] In conclusion, while various implementations and examples have been described herein, they are provided by way of illustration and example, and many potential modifications will be apparent to the skilled person having regard to the specific circumstances of any given implementation. Unless the context clearly indicates to the contrary, it is specifically disclosed that the features of any independent claim and / or its associated dependent claims may be combined with the features of any other independent claim and / or any other dependent claims irrespective of whether such a combination is explicitly recited in the claims. Accordingly, the scope of the present case should be determined from the appended claims and their equivalents.

[0145] REFERENCES

[0146] [1] Benjamin O Asamoah et al. 2019. A prototype of a portable optical sensor for the detection of transparent and translucent microplastics in freshwater. Chemosphere 231 (2019), 161-167.

[0147] [2] Adam Burgess et al. 2009. Embodying uncertainty? Understanding heightened risk perception of drink ‘spiking’. The British Journal of criminology 49, 6 (2009), 848-862.

[0148] [3] Doyeli Chakraborty and JB Jeeva. 2019. Detection of Heart Rate Using Reflectance Mode Photoplethysmography. In 2019 Innovations in Power and Advanced Com puting Technologies (i-PACT), Vol. 1. IEEE, 1-3.

[0149] [4] Justin Chan et al. 2022. Testing a Drop of Liquid Using Smartphone LiDAR. Proceedings of the IMWUT 2022 6, 1 (2022), 1-27.

[0150] [5] Ashutosh Dhekne et al. 2018. Liquid: A wireless liquid identifier. In Proceedings of ACM MobiSys 2018. ACM, 442-454.

[0151] [6] Milton Friedman. 1937. The use of ranks to avoid the assumption of normality implicit in the analysis of variance. Journal of the american statistical association 32, 200 (1937), 675-701.

[0152] [7] Yongzhi Huang et al. 2021. Lili: liquor quality monitoring based on light signals. In Proceedings of the MobiCom 2021. 256-268.

[0153] [8] Yongzhi Huang et al. 2021. A Portable and Convenient System for Unknown Liquid Identification with Smartphone Vibration. IEEE TMC (2021 ). [9] Yongzhi Huang et al. 2021. Vi-liquid: unknown liquid identification with your smartphone vibration. In Proceedings of MobiCom 2021.

[0154]

[0010] William H Kruskal and W Allen Wallis. 1952. Use of ranks in one-criterion variance analysis. American Statistical Journal 47 , 260 (1952), 583-621.

[0155]

[0011] Narara Leesakul et al. 2013. A new screening method for flunitrazepam in vodka and tequila by fluorescence spectroscopy. Luminescence (2013).

[0156]

[0012] Mark Mirtchouk et al. 2016. Automated estimation of food type and amount consumed from body-worn audio and motion sensors. In Proceedings of UbiComp 2016. 451-462.

[0013] Jorge Saiz et al. 2014. Electrophoretic fingerprinting of benzodiazepine tablets in spike drinks. Electrophoresis 35, 21-22 (2014).

[0157]

[0014] Fei Shang et al. 2022. LiqRay: non-invasive and fine-grained liquid recognition system. In Proceedings of MobiCom 2022. 296-309.

[0158]

[0015] Laura Sheard. 2011. ‘Anything could have happened’: Women, the night-time economy, alcohol and drink spiking. Sociology 45, 4 (2011).

[0159]

[0016] Suzanne C Swan et al. 2017. Just a dare or unaware? Outcomes and motives of drugging (“drink spiking”) among students at three college campuses. Psychology of violence 7, 2 (2017), 253.

[0160]

[0017] Rebecca Webb et al. 2007. A rapid CZE method for the analysis of benzodiazepines in spiked beverages. Electrophoresis 28, 19 (2007).

[0161]

[0018] Frank Wilcoxon. 1992. Individual comparisons by ranking methods. In Break throughs in statistics. Springer, 196-202.

[0162]

[0019] Yanni Yang et al. 2022. HearLiquid: Non-intrusive Liquid Fraud Detection Using Commodity Acoustic Devices. IEEE loT Journal (2022).

[0163]

[0020] Zhigang Yin et al. 2023. HIPPO: Pervasive Hand-Grip Estimation from Everyday Interactions. Proceedings of the ACM IMWUT 6, 4, 1-30.

[0164]

[0021] Shichao Yue and Dina Katabi. 2019. Liquid Testing with Your Smartphone. In Proceedings of ACM MobiSys 2019. ACM, 275-286.

[0165]

[0022] Agustin Zuniga et al. 2021. Ripe or Rotten? Low-Cost Produce Quality Estimation Using Reflective Green

[0166]

[0023] Zhigang Yin, Mohan Liyanage, Agustin Zuniga, Petteri Nurmi, and Huber Flores. 2023. Hedgehog: Detecting Drink Spiking on Wearables. In The 24th International Workshop on Mobile Computing Systems and Applications (HotMobile ’23), February 22-23, 2023, Newport Beach, CA, USA. ACM, New York, NY, USA, 7 pages, https: / / doi.org / 10.1145 / 3572864.3580336

Claims

Claims1. A wearable sensing device for determining whether or not a drink has been spiked, the sensing device including a sensor arrangement configured to be placed near a container including the drink, the sensor arrangement including: at least one light source to transmit light into the container; and at least one photoreceptor to output a received light intensity signal indicative of light which has been transmitted into the container by the light source and reflected out of the container; at least one condition sensor to detect conditions of the sensing device and its environment; the sensing device further including a processor for processing signals from the photoreceptor and the condition sensor, wherein the processor is arranged process the received light intensity signal and signals from the condition sensor to determine whether or not the drink has been spiked.

2. The wearable sensing device of claim 1 , wherein the condition sensor is a motion detector for detecting the motion of the sensing device and wherein the processor is arranged to produce a motion- compensated signal from the received light intensity signal, whereby noise in the received light signal due to motion of the sensing device is reduced.

3. The wearable sensing device of claim 1 or claim 2, wherein the condition sensor is a luminosity detector for detecting background light levels in the region of the sensing device and wherein the processor is arranged to produce a background light-compensated signal from the received light intensity signal or the motion-compensated signal, whereby noise in the received light signal due to background light in the region of the sensing device is reduced.

4. The wearable sensing device of claim 3, wherein the processor is arranged to produce the motion- compensated signal and the background light-compensated signal using a convolution filter.

5. The wearable sensing device of claim 3 or claim 4, wherein the processor is arranged to produce a feature-extracted signal using an autoencoder to extract features characteristic of a drink from the received light intensity signal, or motion-compensated signal or background light-compensated signal.

6. The wearable sensing device of claim 5, wherein the processor is arranged to model the output of the feature-extracted signal from the autoencoder to determine the extent of drink spiking.

7. The wearable sensing device of claim 6, wherein the processor is arranged to use a convolutional neural network to model the output of the autoencoder.

8. The wearable sensing device of any preceding claim, wherein the condition sensor is a proximity detector for detecting when the sensing device is in the vicinity of a drink container and wherein theprocessor is arranged to be activated when the proximity detector detects that the sensing device is in the vicinity of a drink container.

9. The wearable sensing device of any preceding claim, wherein the light source comprises a red, green, blue light emitting diode (RGB LED) configured to emit red, green or blue light and combinations thereof.

10. The wearable sensing device of any preceding claim, wherein first and second light sources are provided on opposing sides of the photoreceptor.

11. The wearable sensing device of any preceding claim, further comprising a pressure sensor to determine a contact force between the sensor and the container.

12. The wearable sensing device of any preceding claim, further comprising an accelerometer to determine the orientation in space of the wearable sensing device.

13. The wearable sensing device of any preceding claim, wherein the wearable sensing device is a smart ring, wherein the light source is configured to transmit light in a direction radially outwards from the smart ring.

14. The wearable sensing device of claim 13, wherein the smart ring is configured to be worn on a user’s finger with the sensor located in the front of the hand for the sensor to make contact with the container when a user takes hold of the container.

15. The wearable sensing device of any preceding claim, further comprising a communications facility to transmit data on the intensity of the light received by the photoreceptor to an additional device for processing, optionally wherein said additional device is a smartphone and / or a remote server.

16. A method of operating a wearable sensing device for determining whether or not a drink has been spiked, the sensing device including a light source, a photoreceptor and an condition sensor, the method comprising holding the sensing device near a container including the drink; transmitting light from the light source into the container; receiving, by the photoreceptor, light which has been transmitted into the container and reflected out of the container; receiving, by the condition sensor a signal indicative of the environment of the sensing device, processing the signals from the photoreceptor and the condition sensor to determine whether or not the drink has been spiked.

17. The method of claim 16, further comprising performing measurements on the drink when the drink is initially acquired and / or by retrieving reference characteristics for a known type of drink from a database.

Images