Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
AN APPARATUS AND METHOD
Document Type and Number:
WIPO Patent Application WO/2018/220052
Kind Code:
A1
Abstract:
There is disclosed an apparatus and a method for calculating a biomarker value. The apparatus includes an analysis module configured to obtain a heartbeat trace of a subject; determine at least one parameter describing the heartbeat trace, and; calculate at least one output biomarker value using at least one Artificial Neural Network ("ANN"). An input of the ANN includes the at least one parameter. An output of the ANN includes the at least one output biomarker value.

Inventors:
WOOD RICHARD JOHN (GB)
WOOD DOMINIC (GB)
BIBBINGS KATHERINE (GB)
BENNETT ALEXANDER (GB)
Application Number:
PCT/EP2018/064256
Publication Date:
December 06, 2018
Filing Date:
May 30, 2018
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
BIOEPIC LTD (GB)
WOOD RICHARD JOHN (GB)
WOOD DOMINIC (GB)
BIBBINGS KATHERINE (GB)
BENNETT ALEXANDER (GB)
International Classes:
A61B5/00; A61B5/021; A61B5/024; A61B5/145; G06N3/04; G16H50/50
Domestic Patent References:
WO2000022408A22000-04-20
WO2015050929A12015-04-09
Foreign References:
EP3087915A12016-11-02
US20130267796A12013-10-10
US20170112395A12017-04-27
Other References:
HOSSEINI H G ET AL: "The comparison of different feed forward neural network architectures for ECG signal diagnosis", MEDICAL ENGINEERING & PHYSICS, BUTTERWORTH-HEINEMANN, GB, vol. 28, no. 4, 23 August 2005 (2005-08-23), pages 372 - 378, XP028074008, ISSN: 1350-4533, [retrieved on 20060501], DOI: 10.1016/J.MEDENGPHY.2005.06.006
ROHAN BANERJEE ET AL: "Noise cleaning and Gaussian modeling of smart phone photoplethysmogram to improve blood pressure estimation", 2015 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), 19 April 2015 (2015-04-19), pages 967 - 971, XP055303772, ISBN: 978-1-4673-6997-8, DOI: 10.1109/ICASSP.2015.7178113
Attorney, Agent or Firm:
GRAHAM, Emma et al. (GB)
Download PDF:
Claims:
CLAIMS:

1 . An apparatus for calculating a biomarker value , comprising : an analysis mod ule configured to: obtain a heartbeat trace of a subject; determine at least one parameter describing the heartbeat trace; calculate at least one output biomarker value using at least one Artificial Neural Network ("ANN"); wherein an input of the ANN includes the at least one parameter, and an output of the ANN includes the at least one output biomarker value. 2. An apparatus according to claim 1 , wherein the ANN includes an input layer having a plurality of input neu rons, where the number of input neu rons is eq ual to the number of parameters.

3. An apparatus accord ing to claim 1 or claim 2, wherein the ANN includes an output layer having at least one output neuron , where the number of output neu rons is equal to the number of output biomarker values.

4. An apparatus accord ing to any preceding claim, wherein obtaining the heartbeat trace includes deriving the heartbeat trace from a video of the subject.

5. An apparatus accord ing to claim 4, wherein obtaining the heartbeat trace includes measuring an intensity within a target region of each frame of the video, the intensity as a function of time forming the heartbeat trace.

6. An apparatus accord ing to any preceding claim, wherein obtaining the heartbeat trace includes: performing a Fourier transform of the heartbeat trace to identify the heartbeat freq uency. 7. An apparatus according to any preced ing claim , wherein obtaining the heartbeat trace includes measuring a noise signal in the heartbeat trace, and removing the noise sig nal from the heartbeat trace.

8. An apparatus according to any preceding claim, wherein the obtaining the heartbeat trace includes combining a plu rality of individual heartbeat profiles in the heartbeat trace to form a combined heartbeat profile.

9. An apparatus according to any preceding claim , wherein a plurality of ind ivid ual heartbeat profiles are temporally aligned using a feature of the plurality of the individ ual heartbeat profiles.

10. An apparatus accord ing to any preceding claim , wherein the ANN includes a primary AN N and at least one precursor AN N .

1 1 . An apparatus according to claim 1 0, wherein at least one of the input nu merical values to the primary ANN is an output from one of the at least one precursor ANN .

12. An apparatus according to any preceding claim, wherein the input of the ANN includes at least one extrinsic value for the subject.

13. An apparatus according to any preceding claim , wherein determining at least one parameter describing the heartbeat trace includes determining a functional form of the super pulse, and wherein the input of the ANN includes at least one parameter of the functional form .

14. An apparatus according to claim 13, wherein the functional form includes a plurality of Gaussian fu nctions.

15. A method of calculating a biomarker value , comprising : obtaining a heartbeat trace of a subject; determining at least one parameter describing the heartbeat trace ; calculating at least one output biomarker value using at least one Artificial Neural Network ("ANN"); wherein an input of the AN N includes the at least one parameter, and an output of the ANN includes the at least one output biomarker value . 1 6. A computer program comprising instructions which , when the prog ram is executed by a computer, cause the computer to carry out the method of claim 15.

Description:
AN APPARA TUS AND METHOD

This application claims priority from GB1708591.1 filed 30 May 2017, the contents and elements of which are herein incorporated by reference for all purposes. Field of the invention

The present invention relates to an apparatus and method for calcu lating a biomarker value, and more particularly to an apparatus and method for calculating a biomarker value using an artificial neu ral network.

Background Diabetes mellitus (DM), of which type 2 d iabetes mellitus (T2DM) represents 85-95% of cases of diabetes in ad ults, has increased d ramatically to pandemic proportions. T2DM affected 450 million adults in 2014, approximately 8.5% of the world population and is predicted to rise 1 1 .6% by 2025. (Press 201 6)

Increased insulin resistance is a biomarker of T2DM subjects. Impaired g lucose tolerance marks the progression between normal g lucose tolerance and diabetes (Lillioja, Mott et al . 1993), (Reaven 1 988).

Treatment goals for T2DM are to eliminate symptoms, to maintain a normal quality of life and work capability, to prevent the occurrence of acute metabolic d isorder, and to prevent and delay the occurrence and development of chronic complications. Therefore, diabetes treatment is lifelong process. I n addition to insu lin and hypoglycaemic d rugs, diet is the basis for the treatment of diabetes. In all treatments, blood glucose level monitoring is important.

Most of methods for blood glucose level measu rement are invasive . Sample blood from patients is taken and the blood g lucose level is measured , typically by glucose oxidase (GOx) method . Venous whole blood , plasma or serum glucose is tested in hospitals. Capillary whole blood g lucose can be checked on a portable device operated by a patient. To maintain the glucose at a desired level, the blood glucose has to be tested several times per day, including at least before and after the three meals and before bed . If a patient suffers noctu rnal hypoglycaemia, additional testing is needed . All of these methods accurately obtain a blood glucose level, but there are problems and limitations. First, it is very painful to take blood samples mu ltiple times per day. Second , it is costly to use glucose oxidase reagents or test strips in hospital or at home, which presents a significant financial burden to patients. In addition, self-testing at home may lead to blood contamination and bacterial infection.

Therefore development of a non-invasive type of blood glucose monitoring technologies and devices has been a long-term goal of many research institutions and companies. It is an object of the present invention to provide an improved apparatus and method for non-invasive blood glucose monitoring .

Summary of the invention

Accord ing to a first aspect of the invention, an apparatus for calculating a biomarker value is provided , comprising : an analysis module configu red to: obtain a heartbeat trace of a subject; determine at least one parameter describing the heartbeat trace; calculate at least one output biomarker value using at least one Artificial Neural Network ("ANN"); wherein an input of the AN N includes the at least one parameter, and an output of the ANN includes the at least one output biomarker value.

In the context of the present invention, the apparatus may be a mobile device. The mobile device may be a mobile phone, a so-called smart watch, a wearable device such as glasses or contact lens, a computer peripheral (which may be configured for wired/wireless connection to a computer), a tablet, a computer or any other suitable mobile device . The mobile device may be a substantially ded icated biomarker measurement device. For example, the mobile device may be a blood glucose monitor, i .e . a ded icated device for monitoring blood g lucose level. The device may be capable of analysing video, either obtained locally, or su pplied as data to the mobile device. Furthermore, the mobile device may include the means by which to obtain the video. For example, the mobile device may include a camera for obtaining video.

The mobile device may also be capable of recording the biomarker values obtained according to the present invention. The biomarkers may be stored on the device and/or sent to a remote location . I n other words the device may be used to track and record biological and

physiological parameters.

The mobile device may be configu red for external data connection . For example , the mobile device may have a network connection means, for example a wifi or "Bluetooth" wireless connection . The mobile device may be configured to send the biomarker measurements to a remote device, for example a remote server, or another mobile device or computer belonging to the user. Advantageously, the ANN includes an input layer having a plu rality of input neurons, where the number of input neurons is equal to the number of parameters.

Conveniently, the AN N includes an output layer having at least one output neu ron, where the number of output neurons is equal to the number of output biomarker values. Preferably, obtaining the heartbeat trace includes deriving the heartbeat trace from a video of the subject.

Advantageously, obtaining the heartbeat trace includes measu ring an intensity within a target region of each frame of the video, the intensity as a function of time forming the heartbeat trace . Conveniently, obtaining the heartbeat trace includes: performing a Fourier transform of the heartbeat trace to identify the heartbeat frequency.

Preferably, obtaining the heartbeat trace includes measuring a noise sig nal in the heartbeat trace, and removing the noise signal from the heartbeat trace.

Advantageously, the obtaining the heartbeat trace includes combining a plu rality of individual heartbeat profiles in the heartbeat trace to form a combined heartbeat profile.

Conveniently, a plurality of individ ual heartbeat profiles are temporally aligned using a feature of the plu rality of the individ ual heartbeat profiles.

Preferably, the ANN includes a primary ANN and at least one precursor ANN .

Advantageously, at least one of the input numerical values to the primary ANN is an output from one of the at least one precursor ANN .

Conveniently, the input of the ANN includes at least one extrinsic value for the su bject.

Preferably, determining at least one parameter describing the heartbeat trace includes determining a fu nctional form of the super pulse , and wherein the input of the AN N includes at least one parameter of the functional form . Advantageously, the functional form includes a plurality of Gaussian functions.

Accord ing to a second aspect of the present invention, a method of calculating a biomarker value is provided , comprising : obtaining a heartbeat trace of a subject; determining at least one parameter describing the heartbeat trace; calculating at least one output biomarker value using at least one Artificial Neural Network ("ANN"); wherein an input of the ANN includes the at least one parameter, and an output of the AN N includes the at least one output biomarker value.

Accord ing to a third aspect, a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the second aspect.

Of cou rse, the features recited in respect of each particular aspect can be readily applied to the other aspects.

The present invention relates to a mammalian health and wellbeing apparatus that implements a method utilising a combination of technolog ies and systems to establish a non-invasive biomarker value measurement.

More particu larly the apparatus permits a measurement of mammalian homeostasis, and more particularly a biomarker measure of human body regu latory mechanisms.

Homeostasis is su pported by two main functions: the autonomic nervous system and endothelial functions. The autonomic nervous system (ANS) is an extensive nervous network whose main role is to reg ulate the internal environment and body functions by controlling homeostasis which includes haemodynamics, blood pressure, heart rate , blood g lucose level, sweating and visceral fu nctions (Kalopita , Liatis et al. 2014). The ANS acts throug h a balance of stimulation or inhibition of two main components— the sympathetic and parasympathetic nervous systems. Sympathetic and parasympathetic branches act via neurotransmitters and receptors activation.

Endothelial function is related to the ability of the blood vessels to d ilate or constrict when necessary. Endothelial dysfunction can be characterised as reduced bio-availability of N itric Oxide (NO), which plays many roles in maintaining vascular health. Hence, endothelial dysfunction is defined as an impairment of endothelium dependent vasodilation. Endothelial dysfunction gives rise to vascu lar restriction or arterial stiffness which can be measured by a number of techniq ues including flow mediated dilatation using laser Doppler or u ltrasound probes.

A method for monitoring cardiovascular events and peripheral circu lation is throug h localised reflective video analysis (LRVA). LRVA uses reflected red light to measu re relative blood volume in peripheral mammalian tissue such as the fingertip, toe , ear lobe or inner ear. LRVA waveforms are characteristics of blood movement in cutaneous vessels and can be used to identify synchronous depolarization of cardiovascular tissue. The fundamental frequency of the LRVA waveform, typically around 1 Hz representing a heart rate of 60 bpm . Lower frequency components such as respiratory, thermoregulatory and sympathetic nervous system effects are also contained within the LRVA signal. Arterial stiffness, ind icative of endothelial dysfunction , may also be measurable from calcu lations made using the LRVA waveform analysis.

The apparatus/method of the present invention relates to the non-invasive measurement of a biomarker value . For example, a measurement of a blood g lucose level in a human subject. The subject may be a healthy individuals, or an individual that exhibits on-set of or near onset of Type I I Diabetes Myelitis and ind ividuals diag nosed with Type I Diabetes Myelitis. The invention is a means of providing simple tests to identify problems which would not normally result in the need for clinical diag nosis and in this respect are part of the wellness assessment of individ uals.

For many su bjects, it may be inappropriate to undertake ambulatory time-of-day blood g lucose level measu rements to determine blood g lucose levels. The present invention may be particularly usefu l for subjects not previously diagnosed with the symptoms of Type I I diabetes, for example. The present invention allows for a cost-effective biomarker value measurement platform installed on a cell-phone or other such mobile device, or installed on a remote server system to perform rapid biomarker value measurement.

The data so recorded may also be remotely accessed by the subject or a competent person, such as a physician , to review the biomarker value measu rements.

Summary of the figures So that the invention may be more readily understood , and so that further features thereof may be appreciated , embodiments of the invention will now be described by way of example with reference to the accompanying drawings in which:

Fig ure 1 is a schematic view of a mobile device accord ing to an embod iment of the present invention; Fig ure 2 is a schematic view of the steps of involved in deriving input parameters for an ANN in accordance with the present invention;

Fig ure 3A illustrates demonstrates the selection of a target region; Figu re 3B also illustrates demonstrates the selection of a target region; Fig ure 4 illustrates a heartbeat trace ; Fig ure 5 illustrates a super pulse; Fig ure 6 illustrates an ANN in accordance with the present invention; Fig ure 7 illustrates another AN N in accordance with the present invention; Fig ure 8 illustrates the efficacy of the present invention for a first biomarker; Fig ure 9 also illustrates the efficacy of the present invention for the first biomarker; Fig ure 10 illustrates the efficacy of the present invention for a second biomarker, and ; Figu re 1 1 also illustrates the efficacy of the present invention for the second biomarker. Detailed description of the invention

Turning now to consider Figu re 1 in more detail, a schematic illustration of a mobile device 1 according to an embodiment of the present invention is shown. The mobile device 1 includes a camera 2 , storage means 3, an analysis modu le 4, and a memory mod ule 5.

The mobile device 1 may be a mobile phone, a smartphone , a smart watch, a wearable device such as glasses or contact lens, a computer peripheral (which may be configu red for wired/wireless connection to a computer), a tablet, a computer or any other suitable mobile device. The mobile device 1 may be a dedicated biomarker value measurement device. For example , the mobile device may be a blood glucose monitor, i.e . a dedicated device for monitoring blood g lucose level.

The camera 2 is capable of recording video into a video file. The video file is a continuous sequence of images of a scene that has been viewed by the camera 2. The camera 2 may create video files, which may be stored on the mobile device 1 in the storage means 3. The storage means may be RAM or a hard drive , for example . The video file may be sent from the mobile device 1 to a remote server.

The camera 2 may have a frame rate of greater than 24 Hz, preferably in the range 24 to 60 Hz, more preferably in the range of 30 Hz to 60 Hz. The mobile device may be configured to limit the frame rate to a maximum of 60Hz. The frame rate used is a compromise between data rate and noise level. A hig her frame rate leads to higher data rate, which of course the mobile device must be able to handle (perhaps temporarily). The higher data rate gives more data to analyse at a better resolution, however the noise also increases within each frame for shorter frames (hig her frame rate).

The camera 2 may create video files having a so-called Red-Green-Blue ("RGB") three-channel output. The video files may be formed from a seq uence of images , each image being formed from an array of pixel values. Each pixel value may be an RGB value representing the colour of the respective pixel in that frame. Other video formats may be used , for example HVC format. The videos may be converted from one format to another, for example from RGB to HVC format. The storage means 3 may be hard drive or a RAM module , for example . The video files may be temporarily stored on the storage means 3. From the storage means 3 , the video files are accessible for processing by the analysis mod ule 4 (see below).

The analysis module 4 is configured to perform the analysis of a video file in order to calculate the value of a biomarker for the subject appearing in the video. The computer instructions that dictate the operation of the analysis modu le 4 may be stored in the memory module 5. The instructions may be in the form of an application or "app", which can be loaded on to the mobile device 1 . Such instructions form an embodiment of the present invention .

The camera 2 records a video of an area of skin of a subject. The video forms a sequence of frames, having a particular frame rate. For example, the frame rate may be 30 frames-per- second ("fps")), which may be typical for cameras fitted to mobile telephones.

The camera 2 may be a front or back camera of a smartphone or device, for example. The camera may be a stationary mounted video camera.

A video is thus available for processing . The video shows an area of skin of a subject, for example an earlobe or a fingertip. The video may be obtained d irectly from the camera 2 generally immed iately preceding the the processing described herein. Alternatively, the video may be previously recorded , either by the mobile device , or via an entirely u nrelated and/or distinct device. The availability of the video or the relevant data derived from the video (e.g . the heartbeat trace - see below) is what is required for su bsequent processing to measure the biomarker value. A series of steps may then be performed on the video and downstream data derived therefrom. These steps are illustrated in Fig ure 2. The arrow 6 is intended to illustrate the general direction of processing . However, the order of the steps illustrated in Figure 2 should not be considered limiting .

In general , the steps are: · a target area identification step 7 ;

• an intensity calculation step 8; • a frame intensity calcu lation step 9;

• a heart rate/breathing rate determination step 10;

• a noise removal step 1 1 ;

• a heartbeat profile combination step 12, and ; · a su per-pulse formation step 1 3.

All types of camera sensor create a different distortion on the images they produce. This is primarily caused by the geometry of the camera sensor and lens(es). For camera sensors, particularly those in mobile telephones, it has been found that the sensitivity to light across the focal plane of the optical sensor is not flat. In other words, there is variation in the responsiveness of the sensor across the focal plane to given intensity of incident light. It is therefore advantageous to use pixel data from an area of the sensor in which this distortion imposed by the camera sensor is at a minimum .

Returning to Figure 2, the first step is a target area identification step 7. During the target area identification step 7, a target area of the frames of the video is identified for fu rther processing . The target area may be a sub-area of a total frame area of the video.

The target area contains a plu rality of target pixels. The nu mber of target pixels is lower than the total nu mber of pixels per frame of the video. The target area may have a pre-determined shape and size, but the location of the target area within the frames is determined as the area having the lowest distortion . For example, the target area may be the area of the total frame of 100 x 100 pixels that has the lowest d istortion. The target area may be defined as having alternative predetermined shapes and sizes.

The target area may be defined as a zone within the frame of the video where the change in brightness is the greatest as measured by the first derivative of the average brightness of the pixels in the zone. A plurality of test zones may be assessed , where the plu rality of test zones each have a d ifferent location within the frame. For example, the change in brightness for each test zone may be calculated , and the test zone having the greatest change is brightness may be selected as the target area.

The selection of the target area may be performed during a pre-assessment or calibration period .

The target area may also be defined as a region without colou r saturation in either the RGB or HVC spectra . For example, Figure 3A shows a plot of signal response across the area of a typical camera sensor of a mobile device. The x- and y-axes represent position across a camera sensor. The of responsiveness of the sensor corresponds to the curved line. It will of course be noted that position across the camera sensor is analogous to position across a video frame derived from that camera sensor.

It is apparent that the sensitivity is both higher and flatter across the centre of the sensor. The box 17, for example, signifies a selection of the target area within the camera sensor/video frame.

Figure 3B illustrates another aspect of camera sensitivity. The lines on the plot each indicate the average brightness of pixels within a respective test zone). These lines generally correspond to heartbeat traces. The y-axis indicates position of the respective test zone within the camera sensor/frame in one dimension. The x-axis corresponds to time (sample number). Again, it is clear that there is distortion imparted to the heartbeat trace as a function of position across the sensor. The target area may be selected on the basis of the trace exhibiting minimum of distortion.

The target area is used in the measurement of the video but may be re-assessed for each new video.

In the intensity calculation step 8, for at least each of the target pixels in the target area, an intensity value is calculated . The intensity value may be calculated for every pixel in the frame before selection of the target pixels. The intensity for each pixel may be calculated based on a weighted sum of the R, G, and B values for the respective pixel.

In the frame intensity calculation step 9, a single intensity value for each target area in each frame is calculated. The frame intensity for a particular frame may be the average of the intensity values of the target pixels in the target area of that particular frame. An example plot of frame intensity (y-axis, arbitrary units) as a function of time (x-axis, seconds) is illustrated in Figure 4. Intensity as a function of time corresponds to a heartbeat trace. The heartbeat profiles are apparent as intensity fluctuations (the y-axis) as a function of time (the x-axis). Heartbeat profiles are sequentially present in the heartbeat trace of Figure 4.

For example, a proof of concept data collection trial was undertaken to demonstrate that a finger and facial LRVA using the red, green and blue colour scale was used to measure the characteristics of the systolic and diastolic changes in the video intensity. In this case, the video signal was converted into a frame intensity value as a simple average of the pixel colour values. The computations and memory req uired to complete the LRVA of a Super-Pu lse may not be compatible with some mobile devices. As a result the inventors disclose a system in which the LRVA analysis may be performed off-device, for example on a remote server.

In the heart rate/breathing rate determination step 10 , the heart rate is determined by performing a Fou rier transform of the heartbeat trace in the time domain (for example, using a "fast Fou rier transform" (FFT) algorithm). The resu ltant power spectrum in the freq uency domain can be analysed to identify the highest peak, which may correspond to the heartbeat rate in the heartbeat trace.

Similarly, the breathing rate of the subject may also be identified in the frequency domain power spectrum , and recorded .

Advantageously, the video frame rate is greater than or equal 30 frames-per-second (fps). The total length in time of the video image sequence used for determining the heart rate and respiration rate may be less than 30 seconds, for example.

A measurement of the erraticism of the heart rate, by contrast, may be derived from a video having a greater length . For example, a video having a length of greater than 60 seconds, in order to captu re the desired sig nal, for example.

The erraticism of the heart rate is a measure of the variability of the heart rate . Very low heart rate variability (i.e . low erraticism) may be a marker of stress (or "burnout") where the subject enters an unresponsive state. High heart rate variability (i.e. hig h erraticism) may be a characteristic of excess sympathetic nervous system response, which may u ltimately lead to heart arrhythmia.

For example, the heart and respiration rate sequence is preferably half the length (in time) of the heart rate variability sequence with the heart and respiration rate sig nal analysis being performed at least twice every time the heart rate variability is analysed . It may be desirable to minimise the total length of time of the video to minimise continual variability resulting from undesirable artefacts embedded in the sig nal (i.e. noise), such as those arising from movement and the micro-changes on positioning of the camera that records the video. Noise in the heartbeat trace may also arise from su bject movement, sou rce light flicker and video light loss, for example. The noise in the heartbeat trace cou ld be physiological in orig in, or non-physiolog ical in orig in. Reduction and/or removal of such noise happens in the noise removal step 1 1 . In order to eliminate or reduce noise that is non-physiological in orig in (for example , signal artefacts resu ltant from motion between the su bject and the camera during the recording of the video) it may be advantageous to apply sig nal filtering to the heartbeat trace . Typical time- series filtering techniques may be applied . One such example is identifying a portion of the power spectrum in the frequency domain as correspond ing to a noise component, and removing that noise component in the freq uency domain . An inverse Fourier transform is then performed to arrive at a filtered heartbeat trace with the noise component removed . By applying such heartbeat trace noise removal, and thereby removing unwanted noise components of the heartbeat trace, a more accurate biomarker value may ultimately be calculated .

Noise elimination/reduction can alternatively or add itionally be achieved by reconstructing a "noise base signal" by removing the heart rate and respiration rate frequency peaks from the Fou rier transform of the heartbeat trace, and then reconstructing the noise signal in the time domain by performing an inverse Fourier transform. The noise signal in the time domain can then be removed from the original heartbeat trace to provide an improved heartbeat trace with reduced noise.

In another potential method of noise reduction , a stabilisation portion of the video at the start of a particular video may be excluded from further processing . The stabilisation portion may be used to allow stabilisation of the video. After stabilisation of the video, the stabilisation portion of the video may be discarded . This may be done because, when the video is initially switched on, the video signal may be erratic. After a stabilisation period , the video signal may settle down or stabilise. It has been identified that the stabilisation period may be approximately 5 seconds for example. Thus, the stabilisation portion that is d iscarded may include the initial portion of the video having a length equal to the stabilisation period . Excluding the stabilising portion of the video may reduce noise.

In another potential method of noise reduction, the mobile device may be config ured to use measurements from an accelerometer to restart the processing if motion artefacts exceed a predetermined threshold .

In addition , noise components in the heartbeat trace can also be caused by physiological changes in the subject that occur during the period that the video is recorded . Such physiological changes can manifest as variations in the heartbeat trace. Such noise components may not have their orig ins in noise derived from motion artefacts or harmonics. A pulse trace containing physiological-origin noise of this type may resu lt in lower accuracy biomarker values being measured therefrom . This kind of noise may be mitigated in a heartbeat profile combination step 12 and the super pu lse step 13. When the heartbeat rate has been identified , a next step is to improve the sig nal to noise ratio of the heartbeat profiles in the heartbeat trace. One way to achieve this is to combine the individual heartbeat profiles with one another. This happens du ring the heartbeat profile combination step 12. Using the heartbeat rate , from which a heartbeat period can be calcu lated , the heartbeat trace can be separated temporally into a plurality of heartbeat profiles each having a length of one heartbeat profile (see Figure 4). Each heartbeat profile corresponds to a sing le heartbeat period , and thus includes a single heartbeat.

The heartbeat profiles may be combined to form a single combined heartbeat profile. An example method of combining the heartbeat profiles is to add the profiles together sample by sample. Temporal alignment between the individual heartbeat profiles before adding of aligned samples may be achieved based on alignment of a particular feature or features of the shape of the ind ividual heartbeat profiles. This allows for more accu rate alignment, and thus improved SN R of the combined heartbeat profile relative to the ind ivid ual heartbeat profiles. The combined heartbeat profile intensity values may be baselined to zero, or some other value . The intensity values in the combined heartbeat profile may be scaled .

Once the combined heartbeat profile has been formed , a "super pu lse" is formed d uring a su per-pu lse formation step 13. The su per pu lse is a functional representation of the combined heartbeat profile. One such functional form is a combination of a plurality of Gaussian fu nctions. The combination may be a summation of the plurality of Gaussian functions. The fitting of the Gaussian functions to the combined heartbeat profile may be achieved by applying a decision tree. The decision tree therefore effectively synthesises the combined heartbeat profile.

For example, synthesising nine Gaussian fu nctions, and fitting those to the combined heartbeat profile, results in an advantageous representation for a super pulse . Each Gaussian function can be defined by three parameters - the central value , the width (e.g . the full-width half-maximu m), and the peak height. For example, three Gaussian functions can be described by 9 parameters; five Gaussian fu nctions by 15 parameters; and nine Gaussian functions by 27 parameters, for example. Fig ure 5 shows a super pulse, formed from the summation of five Gaussian fu nctions. The heartbeat profiles that were used to form the combined heartbeat profile are those shown in Fig ure 4. The combined heartbeat profile was used to form the super pulse shown in Figu re 5. Each Gaussian function may correspond to a feature in the combined heartbeat profile . These parameters of the Gaussian functions are determined by using a non-linear optimisation method whereby an objective "cost" function represented by the minimisation of the sum of the squares of the difference between the combined heartbeat profile and a test Gaussian solution is calculated. When the parameters of the test Gaussian solution are determined such that the objective cost function has a minimum value (for example, <0.0001 ) then the Gaussian parameters for the super pulse may be determined. The Gaussian parameters for the super pulse may be contained in a parameters vector. The values in the parameters vector may be converted into an absolute index. The absolute index is a numeric value that uniquely characterises the individual subject, and may change with medication intake and lifestyle. The frame rate of the video file ultimately dictates the original sampling rate of the heartbeat trace and of the combined heartbeat profile.

The functional form of the super pulse is known (by the plurality of Gaussian functions and the corresponding parameters, for example). Accordingly, it is possible to increase the sampling rate of the super pulse, relative to the original sampling rate, to a higher sampling rate during the interpolation step 14. The higher sampling rate allows for a higher temporal resolution in the super pulse than the temporal resolution of the combined heartbeat profile. Accordingly, temporally smaller features and shorter time periods within the super pulse can be identified and utilised.

For example, the original sampling rate may be 30 Hertz and the higher sampling rate may be 1000 Hertz. The samples that are created in the super pulse in the process of increasing the sampling rate (i.e. the samples added between the actual samples from the video) may be interpolated based on the value of the Gaussian function(s) at that point in the super pulse.

The combined heartbeat profile is similar in utility to an individual's fingerprint. In other words, the super pulse is characteristic of an individual. The mobile device may be configured to verify the identity of the individual using the mobile device by comparison of the super pulse to a reference super pulse for a user. Furthermore, the characteristics (for example, the shapes, sizes, and dimensions) of the combined heartbeat profile and its components depend on the value of a great many biomarker values of the subject.

Identifying a functional form that accurately describes the characteristics of the combined heartbeat profile allows for accurate measurements of the characteristics of the combined heartbeat profile and its components via analysis of the super pulse. Such measurements can then be used the calculation of a biomarker value or values.

It is also noted that the characteristics of the combined heartbeat profile, and by extension the characteristics of the super pulse, may be modified by the action of a medication or functional foods, vitamins, or foods in a changed diet and exercise routine and fu rthermore can be used to identify the result of consuming food on a biomarker value. Furthermore , the characteristics can be shown to vary with the su bject's circad ian rhythm .

Where the super pu lse is represented using a plurality of Gaussian functions, the parameters that characterise the su per pu lse may be recorded as a su bject vector. The absolute value of the subject vector may provide a convenient and reproducible single numeric super pu lse value.

The super pulse may also allow acute and chronic changes on vascular health to be recorded and related to other factors such as pulse wave velocity, variable heart rate, stiffness index and flow mediated dilatation and stress and mental fatigue indexes.

The su per pulse also contains valuable information about the orthogonal orientation of red and white blood cells with respect to time. The orthogonal orientation of the blood cells defines the velocity of travel of the cells.

For example , it may be possible to identify the point of the systole in the super pulse. The time of the systole may form an input to the ANN (see below). The systole point may be timestamped , using the super pulse of the present invention .

The super pulse may also be analysed in such a way as to highlight the change in alertness of the subject. This may be achieved , for example, by computing the areas under the super pulse for that part of the super pulse representing the systole and the diastole . Such areas may form inputs to the ANN (see below). The ratio of these areas may change as a d irect result of a su bject u ndertaking a task where the use of a medication , or functional foods, vitamins, or foods in a changed d iet, or exercise. The ratio of the areas may form an input to the AN N (see below). Such example activities may directly affect this ratio. In this way, the subject's response to a particular intervention can be measured and presented as an alertness index. An alertness index may be calculated by a precursor ANN , and input to the primary AN N .

In a fu rther embodiment of the invention, the mobile device may also be used to record the med ications , fu nctional foods, vitamins, or foods consumed on a daily basis by scanning a barcode. In this way, a clear record of consu mption can be made without recou rse to "after the event" food consumption q uestionnaires. When all the aforementioned aspects are combined it is possible to identify the correlation of the power spectral density of the recorded information and utilise that information to correlate with measured insulin resistance. The power spectral density of the video provides another set of parameters, for example parameters describing heartrate variability (erraticism). Accord ing to the present invention, the mobile device is able to non-invasively calcu late a biomarker value of a subject. Examples of such a biomarker are insulin resistance, blood glucose concentration, systolic and diastolic blood pressu re and core or skin surface temperatu re , red blood cell count, histamine levels and/or immu noglobu lins for example . Eq ually, the present invention may be configu red to calculate a plurality of biomarker values.

By way of example, the following d iscussion relates to blood glucose level as the biomarker.

In the present invention, an Artificial Neural Network ("ANN") module is implemented in the analysis step 1 5 (see Fig ure 2). The analysis step is implemented in the analysis module 4 of the mobile device 1 . Alternatively, the analysis step 15 may be implemented on a remote server, remote from the mobile device 1 .

An AN N is a machine learning technique . I n general, an ANN is based upon a plurality of interconnected (in a mathematical sense) computational units (commonly referred to as "neurons"). Generally, the neurons are connected in layers, and signals travel from the first (input) layer, to the last (output) layer. The variables (incoming and outgoing ) and the state of each neuron is a real number, typically chosen to be between 0 and 1 .

Fig ure 6 illustrates an example of a feedforward artificial neu ral network (FNN). An FNN is a category of ANN . The FN N includes an input layer 22 and output layer 23. Between the input layer 22 and the output layer 23 is a sing le hidden layer 24. The input layer 22 includes input neurons 22A; the hidden layer 24 includes hidden neurons 24A, and ; the output layer 23 includes output neurons 23A. There are input neu ral connections 25 between the input neurons 22A and the hidden neurons 24A. There are output neural connections 26 between the hidden neurons 24A and the output neurons 23A. Data flows one way from the input layer 22, through the hidden layer 24, to the output layer 23. I n an FNN , the data only flows this way. An FN N (and an AN N , in general) may have more than one hidden layer 24. The input to the input layer 22 may be an input vector. The input vector may be a plu rality of numeric or Boolean values, each respective value in the input vector being used as an input to a corresponding input neuron 22A in the input layer 22. The output from the output layer 23 may be an output vector. The output vector may be a plurality of output values, each respective output value in the output vector being output from a corresponding output neu ron 23A in the output layer 23. The length of input vector may be d ifferent from the length of the output vector. The length of input vector may be equal to the length of the output vector.

In general, each neu ron 22A, 23A, 24A calculates a weighted sum of all of its inputs. The neuron may also apply a bias. The weighted sum is used as the argu ment of an activation function for the respective neuron . The output of the activation fu nction is used as the output of the neu ron. The output of the neuron may be fed into one or more downstream (i.e. su bsequent layer) neurons through one or more outbound neuron connections from the neuron .

Using an ANN may allow for nonlinear models, non-Gaussian belief distributions and u nknown or only partly known modelling eq uations. Missing knowledge about the system is extracted during the training process of the AN N .

AN Ns have been designed and implemented to solve regression problems. Reg ression involves estimating a mathematical relationship between variables. For example , reg ression may correspond to the process of fitting continuous curves onto noisy data - thus estimating the mathematical form of a relationship between the variable as described by the noisy data.

By using ANNs, relationships between variables can be described that are hig hly nonlinear and may be of arbitrary complexity. The complexity of the relationship that the ANN can describe may be limited by the choice of hyper-parameters of the AN N . The hyper-parameters of an AN N correspond to the descriptors of the ANN itself, rather than the underlying data. For example , the hyper-parameters may include the number of neu rons in the ANN and the number of layers into which those neurons are arranged and the connections therebetween.

The ANN mod ule of the present invention may include a single AN N similar to that shown in Fig ure 6, or may include multiple such ANNs. I n the case of multiple ANNs, the ANN mod ule may include at least one primary ANN , and at least one precursor ANN , in which an output of the precu rsor ANN forms an input to the primary ANN . The ANN modu le may include a primary AN N and a plurality of precursor ANN , in which an output of each precursor ANN forms part of the input to the primary ANN .

Fig ure 7 shows a schematic of ANN module 30 in accordance with the present invention.

The ANN mod ule 30 includes a primary AN N 31 (in this case , an FNN) and two precursor AN Ns 32A, 32B. The ANN mod ule uses input 33 to calcu late output 34. The data flows from input 33 towards output 34 (i .e . generally from left to right in Fig ure 7).

The input 33 may include, for example only, any combination of the following :

• d irect super pulse values, d irectly describing the super pulse, including : o at least one parameter defining the functional form of the super pulse (for example , the Gaussian parameters);

• indirect super pu lse values, derived from the super pulse, including : o a relative time period between two points of the super pulse (for example, the time between systolic and diastolic peaks, which may be defined as the phase difference between systolic and diastolic peaks); o a mag nitude of a point on the super pulse o a ratio of the magnitudes of two points of the super pu lse; o a ratio of the areas beneath components of the super pu lse (for example the area corresponding to the systole and the area corresponds to the diastole); o depth of a d icrotic notch

• measured video values, derived or obtained without direct reference to the super pulse but u ltimately derived from the video, including : o heart rate ; o breathing rate; o heart rate erraticism .

• extrinsic values that describe the su bject, includ ing : o Chronological age o Health status (e.g . d iabetic status) o Heig ht o Weight o Gender o Body Mass Index ("BMI") o Country of residence o Nationality o Ethnicity o Exercise routine o Medication consumed o Functional foods consu med o Vitamins consumed o Sleep periods

The input 33 may form a vector of such values. Some of these values may be determined via a precursor ANN before input to a primary ANN (see below).

The output 34 may include one or more biomarker values. At least one biomarker values may be cardiovascular biomarker. Alternatively or additionally, at least one biomarker value may be haemal biomarker. The one or more biomarker values may include , by way of example only:

• Blood glucose level;

• Blood pressu re (systolic and diastolic);

• Core and/or skin temperatu re ;

• Heart rate;

• Vascu lar age (which may be an indicator of heart health , and not necessarily equal to chronological age);

• Arterial stiffness index;

• Aug mentation index;

• Left ventricle eject time ;

• Flow mediated dilatation ;

• Blood gas saturation;

• Insulin resistance ;

• Stress levels;

• Mental fatig ue levels;

• Histamine levels • Immunoglobulin levels;

• Hormone levels

• Red blood cell count.

• White blood cell count

• Blood clotting characteristics (FVIII & FIV, for example)

• ALT;

• Alkaline phosphatase level;

• Bilirubin (Total);

• Calcium level;

• Carbon dioxide level;

• Chloride level;

• Creatinine level;

• Glucose level;

• Potassium level;

• Phosphate level;

• Sodium level;

• Urea level;

• Nitrogen (BUN) level

• A Biochemistry C Profile, including for example, beta 2 microglobulin, BNP, Serum Electrophoresis (EPS), FSH, IGF-1, Immunofixation - serum or urine, Immunoglobulins (A,G,M), Paraprotein measurement (Densitometry), Progesterone, PSA, Prolactin, Testosterone, TGB, Thyroglobulin/TG auto Ab, Total T3, Transferrin/TIBC, TSH.

The output list above may include values that are measured accurately and used in the training and validation of the ANN module (see below). The precursor ANN 32A, 32B are each used to determine one or more precursor output values. A precursor ANN may be useful where it is known , or suspected , that a particular value may be usefu l to the predictive capability of the primary ANN 31 .

For example, a particular precu rsor ANN may take all or part of the input 33 as an input. The desired precursor output values 32C, 32D from the respective precursor ANN 32A, 32B, may be defined as, for example:

• Breathing volume. This may use as input to the precursor ANN 32A, 32B, for example, breathing rate derived from the heart beat trace and parameters describing the functional form of the super pulse; · Systolic and diastolic blood pressu re (which may be output from the same predcursor

AN N 32A, 32B). This may use as input to the precursor ANN 32A, 32B, for example, parameters describing the functional form of the su per pulse .

• Core temperature. This may form an input to the precursor ANN 32A, 32B, for example, in combination with heart rate parameters. Any particu lar input 33 to the ANN module 30 as a whole may form an input to a precursor AN N 32A, 32B and/or an input to the primary ANN 31 . Multiple precursor ANNs 32A, 32B may share common inputs or have overlapping inputs. Multiple precu rsor AN Ns 32A, 32B may have identical inputs, but d ifferent precursor outputs. A precu rsor output 32C , 32D from a precursor ANN 32A, 32B may form an input to the primary ANN 31 . It is noted that an ANN module not including the precursor-primary ANN structure could , may, using the same inputs, alternatively derive correspond ing relationships as those represented by the precu rsor-primary ANN combination. Indeed , biomarker values can be effectively calculated in this way (i .e. with no precursor ANN).

However, it has been found that the predictive performance (e.g . accuracy) of the ANN modu le 30 in determining the biomarker in the output 34 is further improved by including at least one precursor ANN 32A, 32B. Each precu rsor ANN 32A, 32B is used to derive , as an input to the primary ANN 31 , a precursor output value from the input 33, which is used as an input to the primary ANN 31 . Effectively, the precursor ANN forces the AN N modu le to calculate a particular value that is considered an important variable in biomarker value calcu lation . I n a general sense, for an ANN to be able to calculate an output from a given input, where the AN N has not seen that particu lar input previously, the ANN must have been trained . That is, the ANN must be configu red ("trained") to have the desired predictive behaviour. In very general terms, training means determining the weight and bias for each neuron . I n general, training is achieved by using a training data set in which the input and the desired output is known - this is called supervised training with a labelled data set (see below). The weights and biases of the ANN are changed such that the input of the training data set is converted to the desired output of the training data set as closely as possible. The weights and biases may be optimised by minimising an error fu nction (for example via error backpropagation).

A validation data set, which also includes inputs with desired outputs, may be used to verify and check the performance of the ANN . The validation data set is not used in training the ANN (i .e . determining the weights and biases of the AN N). In other words, the validation data set is separate from the training data set.

Producing the training and validation datasets usually req uires "labelling" of each data set. That means that it has been determined that the output is what is desired from the correspond ing input. Labelling may be time consuming . Training and validation sets are said to contain "matched pairs" of data . Each matched pair is a set of input data with a known or desired output.

For the present invention, the aim for the ANN mod ule is to calculate a biomarker value from non-invasively measured input values. I n order to train the ANN mod ule to achieve that predictive power, the ANN module must be trained to implement a relationship between the non-invasively measured parameters and a desired measu red biomarker value. The desired measured biomarker value may have been measured invasively. The AN N module in particular allows a biomarker value to be calculated from input parameters describing the su per pulse or input parameters derived from analysis of the super pulse (and optionally other inputs listed above, for example). The training data and validation data each comprises a plurality of matched pairs of data. Each matched pair includes a known output and a known input. The known input (e.g . the su per pulse parameters) and the known output (i.e. the desired biomarker value) were measured at generally the same time . The biomarker value may have been measured using conventional methods. For example, a blood g lucose level measu rement could be made using a blood glucose measurement of a blood sample. The super pulse measurements that correspond to that biomarker value are measured generally simultaneously as the biomarker value is measured (or the time at which the sample is taken).

The matched pairs in the training and validation data are thus a set of clinical parameters of a plurality of subjects, a set of LVRA signals from the plurality of subjects, and a set of synchronised biomarker values from the same plurality of su bjects. Other parameters may also be included , for example age, ethnicity, gender and disease status (i .e . extrinsic values).

I n the case of blood g lucose level, synchronised biomarker values can be derived from a g lucometer or from whole blood values where the blood glucose value is determined using , by example , YSI 2900 analyser.

Any precursor ANNs present in the ANN module may be trained separately from the training of the primary AN N . The matched pairs for training the precursor ANN would be the particular inputs used by the precursor ANN . The known output wou ld the particular desired precursor output value. It may be advantageous to train the precursor AN Ns separately from the primary ANN because training AN Ns may be a time consuming/compute-cycle intensive process.

Of course, there may be an u ncertainty in invasively or non-invasively measured biomarker values. This applies to the biomarker values used in the training data.

For example , in the case of glucometer readings of blood glucose level, the read ings represent interstitial fluid readings taken from the dermal layer. Such readings are known to lag behind that of whole blood values in time and value. Su rprising ly, the inventors have d iscovered that the variation in g lucometer blood glucose synchronised measu rements can be up to ±27% with a sample mean error of ±7% over the period of a postprandial test seq uence . As a resu lt, the inventors have discovered that variation in glucometer accuracy may red uce the potential accuracy of an ANN module trained using glucometer measured biomarker values.

To address this issue , the inventors have used "fuzzy clustering" of the glucometer readings. This technique has been implemented to create of a fuzzy cluster of desired biomarker values, which is determined by the error distribution of the g lucometer-measured biomarker values used to train the network. For example, the method by which the biomarker value is measured may be known to have a measu rement standard deviation. For each measurement biomarker value , two further biomarker values may be generated . For example, one fuzzy biomarker value at measured value plus 1 standard deviation , and one fuzzy biomarker value at measured value minus 1 standard deviation. Together the measured value and the fuzzy values form a fuzzy cluster. Each of the measu red value and the further values use the same input. The number of matched pairs of for training and/or validation is thus increased by the use of fuzzy clustering .

In the case of "gold standard" devices such as the YSI 2900, blood glucose error distribution is far smaller, however the principal remains. In such gold standard devices, the mean error in whole blood and plasma measu red blood g lucose level values may range from 1 .8% to 2.5% , for example . As such, if the YSI 2900 were used to generate blood glucose level values for training , a fuzzy cluster with an error spread would range between 1 .8% and 2.5% would result. Clearly, this is a smaller error spread that for glucometer measured blood g lucose level values. Again, however, a fuzzy cluster may be used to define fuzzy biomarker output values and input values (i.e. matched pairs) for the training and validation data.

Thus, the training and validation data for the AN N module may include fuzzy biomarker output values and input values. The fuzzy biomarker output values reflect the u ncertainty in the measurements of those biomarker values. In this way, the inventors have shown that it is possible to determine accurately blood glucose , insulin resistance and blood pressure and other biomarker values for a user from non-invasive LVRA. Furthermore, the inventors have shown that the effect of medications, functional foods, vitamins, or foods consu med on a daily basis or exercise can be reflected in the baseline blood glucose levels and a general improvement in (reduction of) insulin resistance. The aim of the ANN module is that is it can predict the output biomarker value(s) for input data that the ANN module has not seen before, and which has not been used as part of the training process.

If the ANN module d id not produce generalised predictive behaviour, it wou ld correspond to an "over-fitted" ANN . An over fitted ANN is one that can only accu rately predict output data for input data that was used in the training data (i.e. input data that the ANN has seen before). An over fitted ANN would not accurately predict output data for input data that was not used in the training data. Clearly, an over fitted ANN is undesirable.

To demonstrate and measure the efficacy of the ANN module in the present invention , the predictive capability of the ANN modu le is assessed using validation data. The validation data is another set of matched pairs, similar to the training data. However, the matched pairs of the validation data are not included in the training data.

In an example of training of the AN N modu le of the present invention, a Levenberg-Marquardt backpropagation method was used to train the ANN mod ule using a fuzzy training set. The predicted blood glucose measu rements (i.e . output biomarker values) for a validation data set were compared to invasively measured blood glucose measurements in the validation set.

Fig ure 8 shows a Clarke error grid (CEG). I n general, the CEG compares a scatterplot of values from a reference method and a new method u nder test. The CEG provides an illustration of how accurately the new method is able predict the values from the reference method .

In the example of Figure 8, the x-axis is blood glucose level in milligrams per decilitre measured using the reference method ; the y-axis is blood g lucose level in milligrams per decilitre measured using the new method .

The CEG is a consensus accuracy grid which displays reg ions assigned A to E where A is a region where maximum accuracy is required , and B, C, D and E are regions where less accuracy is satisfactory or accuracy is not important or unacceptable . The CEG for blood glucose level self-monitoring was constructed by using the expertise of a large panel of clinicians and is now in widespread use to evaluate the accu racy of blood glucose level measurements made by patients.

The CEG is split into five types of region (as above):

• Region A are those values within +/-15 % of the reference sensor,

• Region B contains points that are outside of +/-15% but would not lead to inappropriate treatment,

• Region C are those points lead ing to unnecessary treatment,

• Region D are those points indicating a potentially dangerous failure to detect hypoglycaemia or hyperg lycaemia , and

• Region E are those points that would confuse treatment of hypog lycaemia for hyperglycaemia and vice versa .

Fig ure 8 also shows the results (the points on the CEG plot) of a method accord ing to an embodiment of the present invention. Figure 8 demonstrates that the ANN module calculates the blood glucose level entirely within clinically acceptable standards for unseen, noninvasive^ measured inputs. The ANN mod ule used to prod uce the resu lts of Fig ure 8 does not use the fuzzy data approach to training described above .

Fig ure 9 is identical to Figu re 8, except that the ANN module used to produce the results of Fig ure 9 does use the fuzzy data approach to training described above. Figure 9 demonstrates that the ANN mod ule calculates the blood glucose level entirely within clinically acceptable standards for u nseen, non-invasively measured inputs (e .g . validation data). Fig ure 9 shows the improved results obtained when using an ANN module trained using fuzzy data points . A significant majority of the points can be seen in the 'A' classification range of the CEG. There is a notably larger spread amongst the results for type 1 diabetics, and this group also contains every data point outside the 'A' classification . One method of testing u pon unseen validation data is through cross-validation of a combined data set. This is a process completed over the course of a number of training rounds. I n each training round , the combined data is partitioned into two sets: the first is the training data set, whilst the second is the testing data set. In a first training round , the ANN is trained using the training data set, before being tested on the testing data set. In a second , subsequent, training round , another partition of the combined dataset is made and the training process is repeated . This allows for a wide range of data subsets to be used whilst retaining the properties of unseen data. Analysing the errors in predictive capability of the AN N modu le from each training round , across all rou nds, provides information on the generalisability of the approach.

A particular form of cross validation that may be used in training the ANN mod ule of the present invention is so-called "leave-one-out cross validation" ("LOOCV"). Here, each testing round uses a single data point as the testing data, and the rest of the data is used for training data. This allows every single data point to be tested as u nseen data, which is advantageous when there is only a limited amou nt of data available.

An alternative form of cross validation that may be used in training the ANN of the present invention is so-called "leave-p-out cross validation" ("LpOCV"). Here, each testing round uses a plurality, p, of data points as the testing data, and the rest of the data is used for training data.

The ANN mod ule used to generate the data shown in Figure 9 was trained by performing LOOCV using a data set derived from healthy, Type I I and Type I subjects using the AN N described above, and the Levenberg-Marq uardt strategy of back-propagation .

The data clearly shows that the system of the present invention is capable of producing 96.4% of the results within the no-clinical risk region of the CEG.

Fig ure 10 illustrates the efficacy of an embodiment of the present invention in which the output biomarker is blood pressure . In particular, Figure 10 is a plot of blood pressure measured according to the present invention using an ANN method described above (on the x-axis) against blood pressure measured using a conventional blood pressure measurement cuff (on the y-axis). The solid circles represent the training data . The open circles represent the results for a test user. The line represents a 1 to 1 linear relationship between the blood pressure calcu lated according to the present invention and blood pressure measured using a cuff.

Fig ure 10 demonstrates that the ANN method for calculating a biomarker value (the blood pressure, in this instance) produces biomarker values that accurately reflect the blood pressure values measured using conventional cuff techniques.

Fig ure 1 1 illustrates a Bland Altman plot also illustrating the efficacy of an embod iment of the present invention in which the output biomarker value is blood pressure . The x-axis of the Bland Altman plot of Figu re 1 1 is the mean of a Mean Arterial blood pressu re measu red according to the present invention using an AN N method and a blood pressure measured using a conventional blood pressure measurement cuff. The y-axis of the Bland Altman plot of Figure 1 1 is the difference (as a percentage) between Mean Arterial Pressu re measured according to the present invention using an ANN method and Mean Arterial Pressure measu red using a conventional blood pressure measurement cuff.

Mean Arterial Blood Pressu re may be defined as Diastolic Pressu re + 1 /3 * (Systolic-Diastolic). The difference between Systolic and Diastolic pressures is called the pulse pressure.

Fig ure 1 1 demonstrates that the ANN method for measu ring the blood pressu re biomarker produces low bias (the ANN biomarker measurements are close to zero on the y-axis) and the mean error is low (the ANN biomarker measurements are clustered between +/- 1 .96 sigma, which corresponds to a 95% confidence interval . In a first potential use of an embodiment of the present invention , subjects who have a need to manage their diet, or who consu me medications such as Metformin in the case of Type I I diabetics, or fu nctional foods, vitamins, or low carbohydrate foods , in a changed diet reg ime as prescribed , may permit the su bject's data to be returned to a central registry. In this way, a population-related benefit of the particular medication or functional foods, vitamins, or foods in a changed d iet can be demonstrated in a pseudo parallel-use environment such that many thousands of participants' data can be assessed . This type of analysis allows the recording of the change in the health of a typical or average user classified by a number of selected parameters such as age, gender, BM I , country, ethnicity, genetics, exercise routine and medication consumed , such that the benefit of the medication or fu nctional foods, vitamins , or foods in a changed d iet can be uneq uivocally demonstrated . This embodiment is particularly valuable for assessing the change in or improvement in subjects on the borderline of being diag nosed with type I I diabetes.

In a second potential use of an embodiment of the present invention, the subject may be an employee of a corporation or organisation in need of med ical or occupational health management. For example, as may be the case with sedentary workers or workers in locations where dust or solvents are prevalent, or particularly for night shift workers be they in general industry or service industries or hospital workers, or military personnel including su bmariners and astronauts. I n such cases, the su bject (e .g . user of mobile device) is an employee, who is advised to undertake a preferred routine such as exercise or to change their diet to include healthy foods, su pplements, vitamins and/or functional foods.

The analysis of the data from the mobile device allows the record ing of the change in the health of the user classified by, and linked to, their medical records includ ing data for a number of selected parameters such as age , gender, BM I , country, ethnicity, exercise routine, medication or functional foods, vitamins, or foods in a changed diet, sleep period , heart rate and its variability (by way of example, not limited here). In this was the benefit of the regimen can be uneq uivocally demonstrated and actions advised to improve and demonstrate the improvements in health of the subject. This embod iment is particularly useful in cohorts where the user is pre-disposed to weight gain or lack of exercise.

In a third potential use of an embod iment of the present invention, the su bject using a mobile device consu mes a product which reduces their increase in insulin resistance, as is the case with people experiencing mild cog nitive impairment. I n this aspect the mobile device measures appropriate physiological parameters (biomarkers), and the mobile device provides an onscreen cognitive test series, specifically related to executive fu nction and memory associated with increase brain g lucose or ketone bod ies. The subject permits their data to be retu rned to a central reg istry such that the popu lation-related benefits of the product which can be medication or a medical food or a food for special medical pu rpose or functional foods, vitamins, or foods in a changed diet can be demonstrated in a pseudo parallel clinical trial. I n this way the costs of acquiring the long-term efficacy data is red uced by many orders of magnitude . I n this context, cognitive confusion or delirium is often associated with urinary tract infections and this invention is particularly usefu l at eliminating those individuals suffering in this way from the overall cohort involved in the cognitive tests.

In a fourth potential use of an embodiment of the present invention, the mobile device is a Bluetooth or similar finger-tip device with an associated product. For example , the associated product may be a blood pressure reducing medication. A daily dose of the medication is combined with the mobile device (which may be an application on a smartphone) which are access enabled with the associated product such that the effectiveness can be monitored remotely by a clinician. I n this way, the effectiveness of the medication can be determined by both the patient, the clinician, and the manufacturer of the medication. I n addition , the manufactu rer gains valuable access to the users' characteristics, irrespective of the point of prescription . The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, expressed in their specific forms or in terms of a means for performing the d isclosed fu nction , or a method or process for obtaining the disclosed resu lts, as appropriate , may, separately, or in any combination of such features, be utilised for realising the invention in d iverse forms thereof.

While the invention has been described in conjunction with the exemplary embod iments described above, many eq uivalent modifications and variations will be apparent to those skilled in the art when given this disclosure. Accordingly, the exemplary embodiments of the invention set forth above are considered to be illustrative and not limiting . Various changes to the described embodiments may be made without departing from the spirit and scope of the invention .

For the avoidance of any doubt, any theoretical explanations provided herein are provided for the purposes of improving the understanding of a reader. The inventors do not wish to be bound by any of these theoretical explanations. Any section head ings used herein are for organizational purposes only and are not to be construed as limiting the su bject matter described .

Throughout this specification , including the claims which follow, unless the context requ ires otherwise, the word "comprise" and "include", and variations such as "comprises" , "comprising", and "including" will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or grou p of integers or steps.

It must be noted that, as used in the specification and the appended claims, the singu lar forms "a ," "an," and "the" include plural referents u nless the context clearly dictates otherwise. Ranges may be expressed herein as from "about" one particular value , and/or to "about" another particular value . When such a range is expressed , another embodiment includes from the one particular value and/or to the other particular value. Similarly, when values are expressed as approximations, by the use of the antecedent "about," it will be u nderstood that the particular value forms another embod iment. The term "about" in relation to a nu merical value is optional and means for example +/- 10%.