An accessory device

Abstract

An accessory device 5, suitable for use with an injection device 3, comprising a releasable, attachment mechanism and one or more sensors capturing sensor data indicative of injection events during us

Description

Technical Field The present disclosure relates to injection devices, and in particular to accessory devices for use with injection devices. Background Injection devices are used for the convenient administration of liquid medicaments. For example, injection devices may be used for providing a single metered dose of a liquid medicament. Such devices may be either single use “disposable” devices in which the device is typically provided with a syringe already installed, and which is not user-replaceable, or “reusable” devices which allow the user to replace the syringe when the medicament has been used. Further, injection devices (disposable as well as reusable ones) may be provided in the form of an “autoinjector” device which generally comprises a firing mechanism arranged to deliver the fluid from the syringe automatically under the force of a drive system, such as a drive spring. To successfully administer a dose with an autoinjector, the user needs to perform a number of steps. For example, the user may need to remove a cap of the device, place the device against the patient’s skin, activate the firing mechanism, wait (for the injection to finish), and remove the device from the skin. Mistakes in performing these steps can mean that the dose is not completely delivered (or is not delivered at all) - for example, when a user removes the device too early from the patient’s skin. Thus, frequently it is desired to provide assurance (e.g. as part of an adherence monitoring effort) that the autoinjector was correctly operated (and thus that the medicament was correctly administered). Providing such assurance can be beneficial in home care settings as well as in clinical trials (e.g. to provide assurance that drugs were taken in accordance with clinical trial requirements). Summary The present disclosure aims to provide new and useful devices, systems and methods for enabling the detection and identification of injection events during usage of an injection device, e.g. to provide assurance that the autoinjector was operated correctly. According to a first aspect of the present invention, there is provided an accessory device for use with an injection device (e.g. a disposable or reusable autoinjector) for delivering a dose of medicament. The accessory device comprises an attachment mechanism for releasably attaching the accessory device to the injection device, and one or more sensors configured to capture sensor data indicative of injection events during usage of the injection device. For example, an injection event may be a cap removal event, a start of dose event, an end of dose event, a hold time event, a safety shroud deployment event, or the like. The term “sensor data indicative of injection events” may mean that each of a number of possible injection events may result in a distinct (e.g. substantially unique) pattern (or “signature”) of the values of the sensor data which can be used to reliably distinguish and identify injection events (these distinct patterns may be specific to the “type” or “model” of the injection device, i.e. different types / models may have different patterns for the same injection event, while different instances of the same type / model may have substantially the same patterns for the same injection event). The one or more sensors comprise a motion sensor for sensing movement of accessory device (and therefore the attached injection device), and / or a sound sensor (e.g. a piezo microphone) for recording sounds generated by the attached injection device (e.g. sounds of mechanical processes performed by the injection device such as clicking, sliding, rattling, snapping or other mechanical sounds). The motion sensor may comprise an accelerometer, and specifically a three axis accelerometer configured to collect acceleration / motion data along three orthogonal axes. Embodiments of the present invention may enable electronically detecting injection events of natively “non-smart” injection devices (e.g. conventional disposable autoinjectors). This in turn allows electronically determining whether the administration of the dose was successful / complete. Embodiments may achieve this by capturing, during use of the injection device (i.e. during an attempt to administer a dose of medicament), as sensor data, (at least) a sound recording and / or motion data. The captured sensor data can be further processed to detect and identify injection events (as described below in detail, this processing of the sensor data may be performed by the accessory device itself or, alternatively by a computing device connected, wirelessly or otherwise, to the accessory device). The proposed accessory device is a reusable and releasable attachment for an injection device and may therefore be particularly well-suited for use with disposable injectors (for which it is neither economical nor environmentally friendly to integrate electronic sensors). In an embodiment, the one or more sensors may further comprise a gyroscope for sensing an orientation of the attached injection device. In this case, the sensor data may comprise corresponding orientation data captured by the gyroscope. The provision of orientation data in addition to sound recordings and / or accelerometer data may further improve the accuracy with which injection events can be detected and identified from the sensor data. In an embodiment, the one or more sensors may be configured to continuously or periodically collect the sensor data (i.e. motion data and / or sound data) so as to produce a time-series, or waveform, of sensor values for each of the one or more sensors. For example, a time-series, or waveform, of sensor values for a particular sensor may comprise a sequence of sensor values captured at different (e.g. equally spaced) points in time (e.g. the time-series of sensor values for a particular sensor may comprise for each time point a single numerical value). For example, the one or more sensors may generate time-series data by sampling at a fixed (i.e. predetermined) rate (e.g. the one or more sensors may sample at a rate between 1 kHz to 100 kHz). The time-period spanned by the time-series may cover the entire injection process. Collection of the sensor data by the one or more sensors may be triggered by a “wake-up” event, which may comprise an interaction of the user with the device, such as pressing a button or the like. The collection of the sensor data may continue until a criterion is fulfilled, e.g. the collection of the sensor data may continue for a fixed amount of time, or until an inactivity of the device is detected (e.g. when the values of the sensor data are below a threshold for a predetermined period of time)). Providing time-series (with high temporal resolution) may further improves the accuracy with which injection events can be detected and identified from the sensor data. For example, injection events of conventional injection devices often result in audible and / or abrupt positional distortions with characteristic temporal patterns. In an embodiment, the accessory device may further comprise a signal pre-processing unit configured to generate down-sampled sensor data by down-sampling at least one of the timeseries and to provide, as output of the one or more sensors, sensor data comprising the down-sampled time-series. For example, when the characteristic patterns of the injection events happen on a time scale which is slower than the sample rate of a particular sensor, it may be advantageous to down-sample the corresponding time-series (e.g. to “trade” temporal resolution for an improved signal-to-noise ratio and reduced data size) before further processing the time-series. More specifically, in an embodiment, the one or more sensors may comprise a high-speed accelerometer that samples at a high rate (e.g. higher than 16 kHz) and generates a corresponding time-series having a high temporal resolution. In this case, the signal pre-processing unit may process this high-resolution time-series to generate a down-sampled time-series which has a lower temporal resolution (e.g. 2 kHz or less, or between 8 and 16 kHz). In an embodiment, the accessory device may further comprise a computing module configured to detect and classify the one or more injection events based on the motion data and / or the sound data. For example, the computing modules may be configured to detect and classify the one or more injection events based on a comparison between the collected motion data and / or sound data and known injection event motion and / or sound data. Alternatively, the computing modules may be configured to detect and classify the one or more injection events by performing the below described computer-implemented method. In an embodiment, the accessory device may further comprise communication module configured to transmit the sensor data to a computing device for detecting and classifying one or more injection events from the transmitted data. For example, the communication module may comprise a Bluetooth transceiver module and / or a wireless network module. According to a second aspect of the present invention, there is provided an assembly comprising an injection device for delivering a dose of medicament, and an accessory device according to the first aspect. The accessory device is releasably attached to the injection device. According to a third aspect of the present invention, there is provided a computer-implemented method of detecting and classifying injection events during usage of an injection device for delivering a dose of medicament. The method comprises receiving, from an accessory device releasably attached to the injection device, data comprising sensor data indicative of the injection events (as described above, “sensor data indicative of injection events” may mean that each of a number of possible injection events may result in a distinct pattern of the sensor data (i.e. a “signature” or “fingerprint” of the injection event)). The accessory device comprises one or more sensors for capturing the sensor data. The method further comprises processing the received data to detect and classify one or more injection events. The sensor data comprises accelerometer data indicative of an acceleration of the attached injection device and / or sound data of the attached injection device. In an embodiment, the sensor data may comprise a time-series of sensor values for each of the one or more sensors of the accessory device (e.g. as described above). In an embodiment, the received data further comprises down-sampled sensor data generated by down-sampling at least one of the time-series (e.g. as described above). In an embodiment, processing the received data comprises using a trained processing module to: process the received data to detect the one or more injection events; generate, for each detected injection event, a score distribution over a set of possible injection event types, and assign, based on the score distributions, one of a corresponding set of injection event labels to each detected injection event. Thus, in broad terms, the processing module receives, as model input, the sensor data (e.g. in the form of one or more time-series), detects (and classifies) distinct injection events in the sensor data and, provides, as model output, data specifying the classified injection events. The processing module may be trained using a training dataset (e.g. using supervised learning methods). The training dataset may comprise a plurality of label training data items. Each training data item may comprise training sensor data and associated labels indicating injection events. The associated labels of the training data items may be obtained by expert annotators. The training dataset may comprise training examples corresponding to correct usage of the injection device as well as examples corresponding to (e.g. intentionally) incorrect usage of the injection device. As noted above, the sensor data patterns of the injection events may differ for different types / models of injection devices, and thus the training dataset may comprise training data item corresponding to the same type / model of injection device. Thus, it is to be understood that, after training, the processing module may be specific to the type / model of the injection device, i.e. the trained processing module may reliably classify injection events of the corresponding injection device but may not very reliable when used with a different type / model of injection device. Notably, the trained processing module may be able to distinguish injection events with a higher accuracy than alternative signal processing approaches, e.g. based on spectral filtering, thresholding, and the like. This means that the proposed method can be used with existing off-the-shelf autoinjectors (i.e. the autoinjectors may not need to be specifically modified, for example to generate particularly distinguishable sounds). This is advantageous because such modifications often require lengthy and costly regulatory (re-)approval of the injection devices. In an embodiment, the set of injection event labels comprises one or more of: a start of dose event, an end of dose event, and a hold time event. In an embodiment, the processing module may comprise a convolutional neural network. In an embodiment, the sensor data may further comprise gyroscope data indicative of an orientation of the attached injection device (e.g. as described above). In an embodiment, the method may further comprise determining, based on the detected and classified injection defects whether the injection device was used correctly. This can be achieved in different ways, for example, the detected events may be compared to a predefined list of expected events (additionally, the order of the detected events and / or the time between the detected events may also be taken into account). In an embodiment, the accessory device may further comprise a communication module for transmitting the sensor data to a mobile computing device, and the method is performed by the mobile computing device (e.g. mobile phone, tablet computer, a laptop, awearable device, or the like). For example, the mobile computing device may belong to the user of the injection device and may run a corresponding application for performing the method. In an embodiment, the accessory device may further comprise a communication module for transmitting the sensor data to a cloud-based application (i.e. an application provided / hosted by a remote computer system (e.g. a data centre) via the Internet), and the method is performed by the cloud-based application. In an embodiment, the accessory device may be an accessory device according to the first aspect. In an embodiment, the accessory device may be an accessory device according to the first aspect further comprising a computing module configured to perform the method. The trained processing module may not only distinguish between injection events with a higher accuracy but also between injection events and non-injection events (e.g. tapping against the injection device by the user). In some embodiments, the trained processing module may also detect and label (one or more) non-injection events. According to a fourth aspect of the present invention, there is provided a system comprising a computing device comprising a computing module, and the accessory device according to the first aspect, wherein the computing module is configured to determine the one or more injection events based on the motion data and / or the sound data received from the accessory device (e.g. by performing the method of the third aspect). In an embodiment, the system may comprise an injection device attached to the accessory device. According to a fifth aspect of the present invention, there is provided a system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform the operations of the method of the third aspect. According to a sixth aspect of the present invention, there is provided one or more computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform the operations of the method of the third aspect. Brief Description of the Drawings Figure 1 shows an illustrateve exampie system for detecting and classifying injection events according to an embodiment. Figure 2 is a block diagram of component of the system of Figure 1. Figure 3 is a flow diagram illustrating an example process of detecting and classifying injection events according to an embodiment. Figure 4 is a block diagram of a training system for training a processing module according to an embodiment. Figures 5 and 6 are experimental results from an example of using the system of Figure 1. Figures 7 and 8 are block diagrams of variations of the medicament delivery system of Figure 1. Detailed Description Figure 1 illustrates an example system 1 for detecting and classifying injection events. The system 1 comprises an injection device 3, an accessory device 5 attached to the injection device 5, a computing device 7 in communication with the accessory device 5 and with a “cloud” storage 9. In the embodiment of Figure 1, the injection device 3 is a conventional disposable autoinjector having a generally elongated housing (in other embodiments, the injection device 3 may be a different type of injection device, e.g. a reusable autoinjector). In general, the conventional autoinjector 3 is configured to automatically deliver a dose of medicament from a syringe (a cartridge, a medicament container, or the like) to a patient upon actuation of a firing mechanism of the autoinjector. To this end, the autoinjector may comprise the syringe, a syringe plunger for driving a bung of the syringe, and the firing mechanism which may comprise a drive system (for driving the syringe plunger), an energy storage mechanism (e.g. a primed spring) and a release mechanism for releasing energy from the energy storage mechanism so that the drive system drives the syringe plunger forwards. To successfully administer a dose with the autoinjector, a user needs to perform a number of steps. It is understood that the specific steps depend on the design details of the autoinjector and on the to-be-delivered medicament. In the following it is assumed that the required steps include removing a safety cap of the autoinjector, placing a forward end of the autoinjector against the patient’s skin, activating the firing mechanism (to start the injection), keeping the autoinjector held against the patient’s skin until the dose is fully expelled plus any additionally required “hold time” (e.g. viscous drugs may require an additional hold time (typically on the order of a few seconds) before the device is to be removed from the injection site), and removing the autoinjector from the skin to allow deployment of a safety shroud. The skilled person will appreciate that these steps are provided as an example only, and that in alternative arrangements, some steps may be omitted, or additional steps may be required. The injection device 3 may include visual and / or audible indicators which indicate specific injection event to a user. For example, the autoinjector may be configured to produce audible clicks at the start and end of dose delivery, and / or to display a bright (yellow) indicator in a viewing window of the device to indicate to the user that the dose has been delivered. However, disposable autoinjectors, like the one shown in Figure 1, do not normally include means for electronically detecting injection events, let alone means for transmitting corresponding digital data to an (external) computing device (e.g. because incorporation of complex electronics (and an associated power source) into a disposable device is not environmentally friendly). Thus, automatic and reliable tracking of a patient’s usage of a disposable autoinjector (e.g. for keeping a digital log) is conventionally not possible. The accessory device 5 is a small (relative to the injection device 3) (battery-powered) electronic device which a user can conveniently attach to (and detach from) an injection device 3 such as the disposable autoinjector. The accessory device 5 comprises an attachment mechanism for releasably attaching the accessory device 5 to the injection device 3. In the embodiment of Figure 1, the accessory device 5 has a sleeve-like shaped housing configured to receive an upper portion of the elongated housing of the autoinjector. The autoinjector may be retained within the accessory device 5 as a result of an interference fit, a bayonet connection or releasable latching mechanism. The skilled person will understand that many different releasable attachment mechanisms may be used to releasably attach the accessory device 5 to the injection device 3. In broad terms, a purpose of the accessory device 5 is to provide electronic sensor(s) for sensing (at least) sounds produced by the injection device and / or motion (such as acceleration) of the accessory device 5 and / or the attached injection device 3 so that the data generated by the sensor(s) can be processed to detect and identify injections events. This enables, for example, verifying and tracking the usage of the attached injection device in a digital manner (e.g. as opposed to relying on manual entries made by the user into a physical journal). As described in detail below, the accessory device 5 is configured to be in wireless communication with a computing device 7 (shown as a mobile phone in Figure 1; in other embodiments the computing device 7 may be a tablet, a laptop, a wearable computing device, a desktop computer, or the like) to enable transmission of data (such as the data generated by the electronic sensor(s)) to the computing device 7. The computing device 7 provides the computing capabilities to process the data from the accessory device 5 to detect and identify the injection events. The computing device 7 (running a corresponding application) may display results of this processing on a display of the computing device 7 for the user’s convenience. The computing device 7 is further in communication (via the Internet) with the cloud storage 9 which may receive and store data from the computing device 7 (e.g. to keep a remote log of the patient’s injection history). Figure 2 shows a block diagram illustrating functional units of the accessory device 5, the computing device 7 and the cloud storage 9. The accessory device 5 comprises one or more sensors configured to capture sensor data. More specifically, the accessory device 5 may comprise a microphone module 51 for detecting sounds (i.e. audio vibrations) from the attached injection device 3. That is, the microphone module 51 may detect sound data. The microphone module 51 may output the sound data as digital time-series data (or just “a time-series” for short), i.e. a sequence of values indicative of the sound amplitude measured by the microphone module 51 at different times, and specifically, over the course of use of the attached injection device 3 to perform an injection. In some embodiments, the sound data collected by the microphone module 51 may be timestamped, i.e. the microphone module 51 may also output a corresponding sequence of times when the values of the sequence were measured. The microphone module 51 may detect the sound data / amplitudes at a predefined sample rate (e.g. at a sample rate between 5 kHz and 50 kHz). In this case, the amplitude values are captured at equally spaced points in time. In some embodiments, the microphone 51 may be configured to detect sounds from the attached injection device 3 by sensing audio vibrations of the air (e.g. the microphone 51 may comprise an electret microphone). In other embodiments, the microphone 51 may be configured to record sounds from the attached injection device 3 by sensing audio vibrations of the injection device 3 (e.g. the microphone 51 may comprise a piezo microphone). Alternatively, the sounds may be measured by detecting vibrations of the housing of the injection device by alternative means such as an audio accelerometer, or an optical camera either alone or in combination with each other or a microphone. Sensing vibrations of the injection device 3 may be advantageous since, in this case, there is a higher chance that the sound detected was generated by the injection device itself. The accessory device 5 further comprises a motion module 53 for sensing (i.e. collecting motion data indicative of) movement of the accessory device 5, and therefore movement of the attached injection device 3. In the exemplary arrangement shown, the motion module 53 comprises an accelerometer configured to collect acceleration data. Similar to the microphone module 51, the motion module 53 may output the measured acceleration data as a digital time-series of acceleration values. The (time-series) data generated by the microphone module 51 and the motion module 53 may collectively be referred to as captured “sensor data”. Each accelerometer value may be a signed value indicative of an acceleration along a measured axis (e.g. in units of G-force (g) or meters per second squared (m / s2)). The motion module 53 may comprise a three-axis accelerometer configured to sense the acceleration of the attached injection device 3 along three axes. In this case, the motion module 53 may output a corresponding time-series for each axis (in addition or alternatively to the individual time-series for each axis, the motion module 53 may output a time-series corresponding to the Euclidean norm of acceleration values). In some embodiments, the accelerometer module 53 may comprise an integrated circuit comprising an accelerometer for sensing the acceleration. The acceleration data may be timestamped. The accelerometer may collect the accelerometer data at a predefined sample rate (which may or may not be different from the sample rate of the microphone module 51). For example, in the embodiment of Figures 1 and 2, the sample rate of the accelerometer is equal or lower than 2 kHz. In other embodiments, the accelerometer may be a “high speed” accelerometer, i.e. the sample rate of the accelerometer may be higher than 2 kHz, for example between 2 kHz and 50 kHz, between 8 kHz and 50 kHz, or between 8 kHz and 16 kHz. Because of the high sample rate, such a high-speed accelerometer is not only sensitive to (i.e. can not only record / detect) “slow”, quasi-unidirectional changes of the acceleration of the attached injection device 3 (e.g. when a cap is removed from the injection device 3) but is also sensitive to (i.e. can record / detect) audible vibrations of the attached injection device 3 (e.g. when the syringe bung abuts with a stop surface at the end of the injection). This means that a “high speed” accelerometer may capture some or all of the vibrations that the microphone module 51 detects. Thus, it is to be understood that, in embodiments where the accelerometer is a “high speed” accelerometer, the accessory device 5 may not comprise the microphone module 51 (i.e. in this cases, the microphone module 51 is optional). In some embodiments, the motion module 53 may comprise a pre-processing unit (not shown in Figure 2) configured to down-sample the (or each) captured / recorded time-series and to provide, as output of the motion module 53, a corresponding down-sampled time-series (i.e. the down-sampled time-series has a lower temporal resolution as the original time-series). For example, in embodiments where a “native” sample rate of the motion module is higher than 16 kHz, the pre-processing unit may down-sample the captured / recorded time-series so that the down-sampled time-series has an effective sample rate of 8 to 16 kHz. In such cases, down-sampling may reduce noise in the collected motion data and the amount of data generated by the motion module 53 while the resulting temporal resolution of the down-sampled time-series is still high enough to resolve the desired signatures of the injection events. The accessory device 5 further comprises a communication module 55 for (wireless) communication with the computing device 7 (e.g. the communication module 55 may comprise Bluetooth transceiver module, near field communication module, wireless networking technology, or the like). The communication module 55 may comprise a transmitter configured to transmit the captured sensor data to the computing device 7. The communication module 55 may further comprise a receiver configured to receive data from the computing device 7 or other external devices. The accessory device 5 further comprises a control module 57 configured to control the capturing / collecting of the sensor data by the sensor(s), i.e. by the microphone module 51 and the motion module 53. More specifically, the control module 57 may interact with the microphone module 51 and the motion module 53 to trigger (i.e. to start) the capturing of the sensor data (i.e. to trigger the start of the time-series). In some embodiments, the control module 57 may trigger the capturing of the sensor data in response to receiving a “wake-up” signal or a “wake-up” input (e.g. the communication module 55 may receive the “wake-up” signal or the “wake-up” signal may be received as a result of manual interaction by the user and with the accessory device 5). Many ways of generating the “wake-up” signal or input exist. In general the “wake up” signal is generated in response to detecting that a usage of the injection device 3 and the accessory device 5 is imminent. As first example, the “wake-up” signal may be generated in response to a user pressing a button on the accessory device 5. As a second example, the “wake-up” signal may be generated by the motion module 53 (e.g. in response to a sudden movement of the accessory device 5, e.g. after a long period of inactivity). As a third example, the “wake-up” signal may be generated in response of the accessory device 5 being attached to the injection device 3 (e.g. the injection device 3 may comprise a conductive label on its housing, and the accessory device 5 may further comprise circuitry to detect the presence of conducting label when the accessory device 5 is installed and may generate the “wake-up” signal in response to detecting the conducting label, or alternatively connection of the accessory device 5 to the injection device 3 may actuate a “wake-up” switch of the accessory device 5). As a fourth example, the computing device 7 may transmit the “wake-up” signal to the accessory device 5 (i.e. the “wake up” signal may be received by the communication module 55), e.g. in response to user input into the computing device (e.g. the user selecting a corresponding app on the computing device 7). The captured sensor data may be stored in a memory of the accessory device 5. The skilled person will understand that many other ways may be used to provide a wake up signal including, but not limited to, detecting removal of the cap of the device due to the removal of the cap pressing a switch to close a circuit or sensors in the label, such as metal strips coming into contact when the injection device is picked up. The control module 57 may also interact with the microphone module 51 and the motion module 53 to stop the capturing / collecting of the sensor data. For example, the control module 57 may also control the microphone module 51 and the motion module 53 to stop the capturing of the sensor data after a predefined time period (e.g. 30 seconds) has passed since the start of the capturing of the sensor data. As another example, the computing device 7 may transmit a corresponding “stop” signal to the accessory device 5 (i.e. the “stop” signal may be received by the communication module 55), e.g. in response to user input into the computing device 7 (e.g. the user confirming to the corresponding app on the computing device 7 that usage of the injection device 3 is completed). The control module 57 further comprises a power source (e.g. a (rechargeable) battery; not shown in Figure 2), e.g. for supplying power to the microphone module 51, the motion module 53, the communication module 55 and the control module 57. The computing device 7 comprises a communication module 71 and a processing module 73. The communication module 71 is configured for (wireless) communication with the accessory device 5 (e.g. the communication module 71 may comprise a Bluetooth transceiver module, near field communication module, wireless networking technology, or the like). More specifically, the communication module 55 is configured to receive the captured sensor data from accessory device 5. The communication module 71 is further configured to transmit processing results of the processing module 73 to the cloud storage 9 via the Internet. In broad terms, the processing module 73 is configured to detect and classify injection events from the sensor data captured by the accessory device 5 (i.e. from the time-series captured by the microphone module 51 and the time-series captured by motion module 53) during usage of the injection device 3 for delivering a dose of medicament. More specifically, the processing module 73 is defined by a plurality of model parameters, and has been trained (i.e. the numerical values of the plurality of model parameters have been determined) using the training system 400 described further below with reference to Figure 4. The processing module 73 may be of any known architecture suitable for performing event detection and classification on time-series data. The processing module 73 receives, as model input, the motion data and the sound data (e.g. in the form of a time-series) captured by the microphone module 51 and the motion module 53. Each time-series may be fed into the processing module 73 as a respective vector comprising a plurality of elements (each element comprising a corresponding numerical value). It is to be understood that the number of elements in the time-series captured by the microphone module 51 and the time-series captured by motion module 53 may be different since they may be sampled at different sample rates (however both time-series span substantially the same time period since the control module 57 synchronises the sensor value capturing by the microphone module 51 and the accelerometer module 53). The processing module 73 processes then the model input (according to the values of the plurality of model parameters) to detect and classify one or more injection events. For example, to this end, the processing module 73 may process the model input to detect the one or more injection events (i.e. identify one or more time periods in the time-series as an injection event), and generate, for each detected injection event, a score distribution over a set of possible injection event types (where a score may be indicative of a predicted likelihood that the respective detected injection event is of the corresponding injection event type). The processing module 73 may then provide the classification result (e.g. a list of the identified injection events, the corresponding score distributions, and the start and end point of the injection event within the time-series). Based on the score distribution for each injection event, one of a corresponding set of injection event labels may be assigned to each defect. For example, the injection event label corresponding to the injection event type having the highest score may be assigned to the respective detected injection event. The processing module 73 may assign one (and only one) label to each element of the time-series (however, this is not a strict requirement, and in some embodiments, the processing module 73 may not assign a label to one or more time-periods within the time-series or may assign more than one label to the one or more time-periods within the time series). In the embodiment of Figures 1 and 2, the set of injection event labels comprises: “ambient” (associated with the injection device being in an idle state), “cap off” (associated with the removal of the cap from injection device 3), “EOD click” (associated with an click produced by the injection device 3 at the end of the dose), “fire” (associated the actuation of the firing mechanism), “handling” (associated with the handling of the injection device 3 or the accessory device 5 before or after an injection), “hold time” (associated continued holding of the injection device against the patient’s skin after the end of the dose), “injection noise” (associated with the injection process), “LOS” (associated with the deployment of a lockout shroud / shield), and “tapping” (associated with a user tapping against the injection device 3). In other embodiments, more, fewer, and / or different injection event labels may be used. The inventors have found that injection events associated with conventional injection devices (in particular disposable autoinjectors) result in distinct sound and motion / acceleration signatures which can be accurately distinguished and classified by the trained processing module 73. Experimental results obtained by the inventors are described further below with reference to Figures 5 and 6. In the embodiment of Figure 1 and 2, sound and motion / acceleration data are processed together by the processing module 73. This may result in a higher accuracy of the classification result compared to processing only sound or only motion / acceleration data (e.g. because some injection events may exhibit a pronounced signature in the sound domain but a weak signature in the acceleration domain while other injection events may exhibit a pronounced signature in the acceleration domain but a weak signature in the sound domain). However, in some embodiments, the processing module 73 may process sound data without motion / acceleration data. Similarly, in other embodiments, the processing module 73 may process motion / acceleration data without sound data (e.g. in embodiments where the accessory device 5 comprises a high-speed accelerometer module (but no microphone module)). The computing device 7 is further configured to process the output of the processing module 73 to determine whether a dose was successfully administered (i.e. whether the user operated the injection device 3 correctly). This can be achieved in many different ways. For example, the detected events may be compared to a predefined list of expected events (i.e. the list of expected sets outs the correct order and the expected time between expected injection events, determined using the timestamps associated with the sounds and / or motion data). Based on the output of the processing module 73 and the determination result of whether a dose was successfully administered, the computing device 7 may update a patient’s specific log stored on the computing device 7. The computing device 7 is further configured to transmit the output of the processing module 73 and / or the determination result of whether a dose was successfully administered to the cloud storage 9 (via the Internet). The cloud storage 9 is configured to store the received data, e.g. to update a patient’s specific log stored in the cloud storage 9. Generally, the patient has access to his data stored on the cloud storage 9. In some embodiments, a health care provider associated with the patient may also have access to the patient’s data stored in cloud storage 9, e.g. for remote patient management and adherence tracking. Further, a drug developing or manufacturing company may have access to the (appropriately anonymised) data stored in the cloud storage 9 (e.g. to provide assurance that drugs were taken in accordance with clinical trial requirements). With reference to Figure 3, an example process 300 of detecting and classifying injection events during usage of an injection device for delivering a dose of medicament is now described. The process 300 may be performed using the system 1 of Figure 1. In an initial step 302, a user (intending to use the autoinjector 5) attaches the accessory device 5 to the disposable autoinjector 3. Next, in response to receiving a “wake up signal”, the accessory device 5 starts capturing / collecting sensor data (i.e. sound and acceleration values) (304). As noted above, the “wake up” signal may be generated in many different ways. In step 306, the user performs the injection using the autoinjector 3. Then, the accessory device transmits the sensor data (captured during step 306, i.e. during the usage of the autoinjector 3) to the computing device 7 for processing (308). In step 310, the processing module 73 (running on the computing device 7) processes the received sensor data to detect and classify injection events. The classification results are further processed to determine whether or not the dose was successfully delivered (312). In step 314, the classification results and / or determination results are transmitted to the cloud storage 9. In step 316, the user detaches the accessory device 5 from the used disposable autoinjector 3 so that the used autoinjector 3 can be disposed of and the accessory device 5 can be used with a fresh disposable autoinjector 3 (step 316 may be performed at any point after step 306). The system 1 and the process 300 described with respect to the Figures 1 to 3 provide an elegant solution to the problem of tracking the usage of injection devices, in particular disposable autoinjectors in a digital and automatic manner. A training system 400 for training the processing module 73 is now described with reference to Figure 4. The skilled person will understand that the training system may implement any suitable machine learning technology, for example a machine learning module comprising a convolutional neural network may be used. The skilled person will also understand that the training system may be implemented on a stand-alone machine learning module whose learning is finalised and transferred to a processing module 73 which does not undertake any further machine learning. This allows multiple identical accessory devices to be produced to be used on multiple auto-injectors, for example for provision to patients or use in clinical studies. Alternatively the processing module 73 may comprise a machine learning module that actively undertakes machine learning. Additionally, the machine learning module may be pretrained before being further trained (or fine-tuned) as described below, i.e. the processing module 73 may be pre-trained for performing event detection and classification on general time-series data. The pre-training, may for example, be used to adapt the processing module to detect events on a different type auto-injector device to that it was originally trained on. In broad terms, the training system 400 can be used to perform a supervised training process to train the processing module 73 to detect and classify injection events associated with a target injection device, i.e. a specific type / model of injection device (e.g. a specific type / model of disposable autoinjector). More specifically, the training system 400 comprises training data 410 comprising a plurality of training items. The training data 410 are used together with a training engine 420 to iteratively update / optimise the values of a plurality of model parameters 430 defining the processing module 73. Each training item comprises a training input 412 and an associated target output 414. Each training input 412 comprises sensor data (e.g. a sound and a motion / acceleration time-series) captured during a (training) usage of an instance of the target injection device. The associated target output 414 represents the expected output of the processing module 73 when processing the corresponding training input 412, i.e. the associated target output 414 comprises annotations in the form of labels indicating detection events in the corresponding training input 412. The target output 414 may be obtained by human experts annotating the training input 412 (for example, the expert annotators may observe (e.g. via a video recording) the user while using the target injection device and thus can determine when (within the timeseries) which steps were performed and they were performed correctly). It is to be understood that for proper model generalisation, the training data 410 may comprise training examples corresponding to correct usage of the target injection device as well as examples corresponding to (e.g. intentionally) incorrect usage of the target injection device. The training is performed over a plurality of training iterations. In each training iteration at least one training input 412 is processed by the processing module 73 to generate a corresponding model output. The model output and the associated target output 414 are fed into the training engine 420 to generate (using known methods such as gradient decent optimisation methods) an update for the model parameters 430. This means that the processing module 73 is iteratively trained so as to produce a model output that matches the target output 414. Because the training data is specific to the target injection device, the processing module 73 can accurately classify injection events of the target injection device but may not be able to accurately classify injection events of other types / models of injection devices. This high specificity of the trained processing module 73 may be advantageous since this reduces the risk that the system 1 can be “tricked”, e.g. through the use of the wrong injection device (e.g. a wrong type / model of the injection device, a knock-off device, or a device that has been tempered with). With reference to Figures 5 and 6, experimental data illustrating the performance of the system 1 are described. Panel 500 of Figure 5 shows an example time-series 501 captured by the microphone module 51 (sample rate about 16 kHz) of the accessory device 5 during operation of an example disposable autoinjector to which the accessory device 5 is attached. Additionally, Figure 5 shows an example time-series 503 captured by an accelerometer of the motion module 53 (sample rate about 2 kHz) of the accessory device 5 during operation of the example disposable autoinjector to which the accessory device 5 is attached. The time-series 501 and 503 are vertically offset for clarity. Several injection events are indicated in panel 500 to illustrate the distinct signatures of the injection events. More specially, reference numerals 505, 507, 509 and 511 respectively indicate the removal of the cap of the autoinjector, the actuation of the firing mechanism, the end of dose and the deployment of the safety shroud. As can be seen, each injection event 505, 507, 509 and 511 results in a distinct response in the audio and acceleration domain. It is further evident, that an injection event can have a strong response in both domains (for example the cap removal 505) while other injections events may have a strong response in one domain and only a weak response in the other (e.g. end of dose 509 exhibits a strong response in the audio domain but only a weak response in the acceleration domain). This illustrates the benefits of the combined use of audio and acceleration data for detection and classification of injection events. Figure 6 illustrates the performance of an example processing module 73 trained with the training system 400 of Figure 4. In particular, Figure 6 shows a “confusion matrix” 601 indicating the accuracy of a performed classification on a verification dataset. As can be seen, the mean accuracy of the model is about 97% indicating an excellent classification performance. Figure 6 shows a similar “confusion matrix” 603 for an alternatively trained machine model that only receives sound data captured by microphone module but no acceleration data. As can be seen, the mean accuracy is reduced compared to 601 but still high (about 89%). In particular, the “confusion matrix” 603 shows that without accelerometer data the model’s classification accuracy of “hold time” events is poor (about 51 %). This further illustrates the benefits of the combined use of audio and acceleration data for detection and classification of injection events. Those skilled in the art will appreciate that various modifications may be made to the above described embodiment without departing from the scope of the present invention. For example, Figures 7 and 8 illustrate variations of the above described system where processing occurs not on the computing device 7 but in a cloud application or on a processor present in the accessory device. More specifically, in the embodiment illustrated in Figure 7, the accessory device 5 is provided as described above but the computing device 7’ does not run the processing module 73. Instead the computing device 7’ transmits the sensor data received from the accessory device 5 to a cloud application 11 (i.e. an application that is hosted on a remote server and that is accessed by the computing device 7’ via the internet). The cloud application implements the processing module 73 which processes the received sensor data as described above. The cloud application 11 may further determine from the output of the processing module 73 whether the dose was successfully administered (as described above for the computing device 7). The cloud application 11 may then provide the output of the processing module 73 and / or the determination result to the cloud storage 9. In the embodiment illustrated in Figure 8, the accessory device 5’ comprises a computing module to implement the processing module 73 directly on the accessory device 5’. This means that the injection event detection and classification are performed directly on the accessory device 5’. Further, the computing module of the accessory device may be further configured to determine from the output of the processing module 73 whether the dose was successfully administered. The accessory device 5’ is configured to transmit the output of the processing module 73 and / or the determination result to the computing device 7” which transmits the data further to the cloud storage 9.

Claims

1. An accessory device for use with an injection device for delivering a dose of medicament, the accessory device comprising:an attachment mechanism for releasably attaching the accessory device to the injection device, andone or more sensors configured to capture sensor data indicative of injection events during usage of the injection device,wherein the one or more sensors comprise a motion sensor configured to collect motion data indicative of movement of the attachment mechanism, and / or a sound sensor configured to collect sound data indicative of sounds generated by the attached injection device.

2. The accessory device of claim 1, wherein the motion sensor comprises an accelerometer and / or the sound sensor comprises a microphone.

3. The accessory device of any preceding claim, wherein the motion sensor comprises a gyroscope configured to collect data indicative of an orientation of the attachment mechanism.

4. The accessory device of any preceding claim, wherein the sensor data comprises a time-series of sensor values for each of the one or more sensors.

5. The accessory device of claim 4, wherein the one or more sensors comprise a preprocessing unit configured to down-sample at least one of the time-series and to provide, as output of the one or more sensors, sensor data comprising the down-sampled time-series.

6. The accessory device of claim 5, wherein the one or more sensors comprise an accelerometer for sensing, at a sample rate of higher than 16 kHz, the acceleration of the attached injection device, and the pre-processing unit configured to down-sample a timeseries of acceleration values to a reduced sample rate between 8 and 16 kHz and to provide, as output of the one or more sensors, sensor data comprising the down-sampled time-series of acceleration values.

7. The accessory device of any preceding claim, further comprising a communicationmodule configured to transmit the sensor data to a computing device for detecting and classifying one or more injection events based on the motion data and / or the sound data.

8. The accessory device of any preceding claim further comprising a computing module configured to detect and classify the one or more injection events based on the motion data and / or the sound data.

9. The accessory device of claim 8 wherein the computing module is configured to detect and classify the one or more injection events based on a comparison between the collected motion data and / or sound data and known injection event motion and / or sound data.

10. The accessory device of claim 8 or 9, wherein the computing module implements a trained machine learning model configured to detect and classify the one or more injection events by:processing the sensor data, comprising the motion and / or sound data, to detect the one or more injection events;generating, for each detected injection event, a score distribution over a set of possible injection event types, andassigning, based on the score distributions, one of a corresponding set of injection event labels to each detected injection event.

11. The accessory device of claim 10 wherein the computing module is further configured to determine, based on the detected and classified injection events whether the injection device was used correctly.

12. The accessory device of claim 10 or 11 wherein the set of possible injection event types comprises one or more of: a start of dose event, an end of dose event and a hold time event.

13. A system comprising:a computing device comprising a computing module; andthe accessory device of claim 7,wherein the computing module is configured to determine the one or more injection events based on the motion data and / or the sound data received from the accessory device.

14. A system according to claim 13, wherein the computing module implements a trained processing module configured to detect and classify the one or more injection events by:processing the sensor data, comprising the motion and / or sound data, to detect the one or more injection events;generating, for each detected injection event, a score distribution over a set of possible injection event types, andassigning, based on the score distributions, one of a corresponding set of injection event labels to each detected injection event.

15. A system according to claim 14 wherein the computing module is further configured to determine, based on the detected and classified injection events whether the injection device was used correctly.

16. A system according to claim 14 or 15 herein the set of possible injection event types comprises one or more of: a start of dose event, an end of dose event and a hold time event.

17. An assembly comprising:an injection device for delivering a dose of medicament, andan accessory device according to any of claims 1 to 12, the accessory device being releasably attached to the injection device.

18. A computer-implemented method of detecting and classifying injection events during usage of an injection device for delivering a dose of medicament, the method comprising:collecting, by an accessory device releasably attached to the injection device, data comprising sensor data indicative of the injection events, the accessory device comprising one or more sensors for capturing the sensor data;processing the received data to detect and classify one or more injection events, wherein the sensor data comprises motion data, collected by a motion sensor of the accessory device and indicative of movement of the accessory device, and / or sound data, collected by a sound sensor of the accessory device and indicative of sounds generated by the attached injection device.

19. The computer-implemented method of claim 18, wherein the sensor data comprises a time-series of sensor values for each of the one or more sensors of the accessory device.

20. The computer-implemented method of claim 17 or 18, wherein processing the received data comprises using a trained processing module to:process the received sensor data to detect the one or more injection events;generate, for each detected injection event, a score distribution over a set of possible injection event types, andassign, based on the score distributions, one of a corresponding set of injection event labels to each detected injection event.

21. The computer-implemented method of claim 20, wherein the set of injection event labels comprises one or more of: a start of dose event, an end of dose event, and a hold time event.

22. The computer-implemented method of claim 20 or 21, wherein the processing module comprises a convolutional neural network.

23. The computer-implemented method of any one of claims 18 to 22 further comprising determining, based on the detected and classified injection events whether the injection device was used correctly.

24. The computer-implemented method of any one of claims 18 to 23, wherein the accessory device further comprises a communication module for transmitting the sensor data to a mobile computing device, and the method is performed by the mobile computing device.

25. The computer-implemented method of any one of claims 18 to 23, wherein the accessory device further comprises a communication module for transmitting the sensor data to a cloud-based application, and the method is performed by the cloud-based application.

26. The computer-implemented method of any one of claims 18 to 23, wherein the accessory device is an accessory device according to any one of claims 1 to 12.

27. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform the operations of the method of any one of claims 18 to 23.

28. One or more computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform the operations of the method of any one of claims 18 to 23.Amendments to the Claims have been filed as follows;CLAIMS:

1. An accessory device for use with an injection device for delivering a dose ofmedicament, the accessory device comprising:an attachment mechanism for releasably attaching the accessory device to the injection device,a plurality of sensors configured to capture sensor data indicative of one or more injection events during usage of the injection device,wherein the plurality of sensors comprise a motion sensor configured to collect motion data indicative of movement of the accessory device, and a sound sensor configured to collect sound data indicative of sounds generated by the attached injection device, anda computing module configured to detect and classify the one or more injection events based on the motion data and the sound data,wherein the computing module implements a trained machine learning modelconfigured to detect and classify the one or more injection events by:processing the sensor data, comprising the motion and sound data, to detect the one or more injection events;generating, for each detected injection event, a score distribution over a set ofpossible injection event types, andassigning, based on the score distributions, one of a corresponding set ofinjection event labels to each detected injection event, andwherein the set of possible injection event types comprises one or more of: a start of dose event, an end of dose event and a hold time event.

2. The accessory device of claim 1, wherein the motion sensor comprises anaccelerometer and / or the sound sensor comprises a microphone.

3. The accessory device of any preceding claim, wherein the motion sensor comprises agyroscope configured to collect data indicative of an orientation of the attachment mechanism.

4. The accessory device of any preceding claim, wherein the sensor data comprises atime-series of sensor values for each of the sensors.

5. The accessory device of claim 4, wherein the sensors comprise a pre-processing unitconfigured to down-sample at least one of the time-series and to provide, as output of the sensors, sensor data comprising the down-sampled time-series.

6. The accessory device of claim 5, wherein the sensors comprise an accelerometer forsensing, at a sample rate of higher than 16 kHz, the acceleration of the attached injection device, and the pre-processing unit configured to down-sample a time-series of acceleration values to a reduced sample rate between 8 and 16 kHz and to provide, as output of the sensors, sensor data comprising the down-sampled time-series of acceleration values.

7. The accessory device of any preceding claim, further comprising a communicationmodule configured to transmit the sensor data to a computing device for detecting and classifying one or more injection events based on the motion data and the sound data.

8. The accessory device of claim 1 wherein the computing module is configured todetect and classify the one or more injection events based on a comparison between the collected motion data and sound data and known injection event motion and sound data.

9. An assembly comprising:an injection device for delivering a dose of medicament, andan accessory device according to any of claims 1 to 8, the accessory device being releasably attached to the injection device.

10. A computer-implemented method of detecting and classifying injection events during usage of an injection device for delivering a dose of medicament, the method comprising:collecting, by an accessory device according to claim 1 releasably attached to the injection device by an attachment mechanism, data comprising sensor data indicative of the injection events, the accessory device comprising sensors for capturing the sensor data;processing the received data to detect and classify one or more injection events, wherein the sensor data comprises motion data, collected by a motion sensor of theaccessory device and indicative of movement of the accessory device, and sound data, collected by a sound sensor of the accessory device and indicative of sounds generated by the attached injection device,wherein processing the received data comprises using a trained processing module to:process the received sensor data to detect the one or more injection events; generate, for each detected injection event, a score distribution over a set of possible injection event types, andassign, based on the score distributions, one of a corresponding set of injection event labels to each detected injection event, andwherein the set of injection event labels comprises one or more of: a start of dose event, an end of dose event, and a hold time event.

11. The computer-implemented method of claim 10, wherein the sensor data comprises a time-series of sensor values for each of the sensors of the accessory device.

12. The computer-implemented method of claim 10, or 11, wherein the processing module comprises a convolutional neural network.

13. The computer-implemented method of any one of claims 10 to 12, wherein the accessory device further comprises a communication module for transmitting the sensor data to a mobile computing device, and the method is performed by the mobile computing device.

14. The computer-implemented method of any one of claims 10 to 12, wherein the accessory device further comprises a communication module for transmitting the sensor data to a cloud-based application, and the method is performed by the cloud-based application.

15. The computer-implemented method of any one of claims 10 to 12, wherein the accessory device is an accessory device according to any one of claims 1 to 8.

16. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform the operations of the method of any one of claims 10 to 12.

17. One or more computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform the operations of the method of any one of claims 10 to 12.s

Images