Artificial intelligence for sound detection from medicament or test device
Techniques for artificial intelligence medicament or test delivery device sound output (MDDSO) detection includes receiving, at a neural network, processed data based on operational data that indicates a time series having a first duration for signals from a vibration detector collected during operation of a medicament device. The neural network classifies the processed data as a first MDDSO or not. The classification data is sent to an output device. The neural network has been trained, prior to receiving the operational data, with first training data and second training data. The first training data indicates time series of ambient sounds, each time series having the first duration. The second training data indicates time series of sound having the first duration during operation of the type of medicament or test device, wherein each time series of the second training data includes sound from the first MDDSO emitted from the device.
1 . A method for artificial intelligence medicament or test delivery device sound output (MDDSO) detection, comprising:
providing a medicament or test delivery device (MDD) comprising an operational component to be manually operated by an operator undergoing training for administering a medicament, with one or more distinct sounds being emitted in response to the operator operating the operational component, and with the medicament or test delivery device corresponding to a first type;
generating by a vibration detector a time series of acoustic data in response to receiving the one or more emitted distinct sounds;
receiving, at a processor configured as a neural network, operational data that includes the time series of acoustic data as having a duration of at least a first duration;
generating automatically on the neural network classification data that classifies the operational data as a first MDDSO or not by inputting processed data based on the operational data into an input layer of the neural network; and
sending the classification data to an output device, with the classification data providing feedback to the operator to indicate whether the first type of medicament or test delivery device was correctly operated when the operational data is classified as the first MDDSO, with the feedback being in terms of at least one of resistance pressure, measure of an amount of medicament delivered, and detect proper forces and sequence on operation of the operational component, and to display a pass or fail result;
wherein prior to receiving the operational data, the neural network has been trained based on the following:
collecting a first training time series corresponding to first training data that indicates a plurality of time series of ambient sounds, each first training time series having the first duration,
collecting a second training time series corresponding to second training data that indicates a plurality of time series of sound having the first duration during operation of the first type of medicament or test delivery device, each second training time series of the second training data includes sound from the first MDDSO emitted from the first type of medicament or test delivery device,
dividing the first duration of the first training data and the second training data into one or more window intervals,
processing each window interval into one or more values for a corresponding number of parameters characterizing sound,
processing acoustic time series data in each window interval to produce one or more values for a corresponding number of parameters characterizing sound in the window interval, and
providing the processed acoustic time series data in each window interval to an input layer of the neural network, and with an output layer of the neural network to produce correct classification data for the one or more values of the processed acoustic time series data provided to the input layer.
2 . The method as recited in claim 1 , wherein:
prior to receiving the operational data, the neural network has been trained further with third training data that indicates a plurality of time series of ambient sound having the first duration during operation of the first type of medicament or test delivery device, wherein each time series of the third training data includes sound from a different second MDDSO emitted from the first type of medicament or test delivery device, wherein the second MDDSO follows in time the first MDDSO during normal operation of the first type of medicament or test delivery device; and
the classification data further classifies the operational data as the first MDDSO or the second MDDSO or neither.
3 . The method as recited in claim 2 , wherein:
prior to receiving the operational data, the neural network has been trained further with fourth training data that indicates a plurality of time series of ambient sound having the first duration during operation of the first type of medicament or test delivery device, wherein each time series of the fourth training data includes sound from a different third MDDSO emitted from the first type of medicament or test delivery device, wherein the third MDDSO follows in time the first MDDSO during normal operation of the first type of medicament or test delivery device; and
the classification data further classifies the operational data as the first MDDSO or the second MDDSO or the third MDDSO or none.
4 . The method as recited in claim 1 , wherein the certain parameter characterizing the sound in each time interval is a power spectrum value for a particular frequency interval.
5 . The method as recited in claim 4 , wherein the power spectrum value for a particular frequency interval is a plurality of power spectrum values for a corresponding plurality of frequency intervals.
6 . The method as recited in claim 5 , wherein the plurality of frequency intervals comprises a number of frequency intervals in a range from about 2 to about 40 frequency intervals.
7 . The method as recited in claim 6 , wherein:
the MDDSO is a click; and
the plurality of frequency intervals spans a frequency range from about 0 Hertz to about 4 kilohertz.
8 . The method as recited in claim 1 , wherein the certain parameter characterizing the sound in each time interval is a value for a particular cepstral coefficient interval.
9 . The method as recited in claim 8 , wherein the value for a particular cepstral coefficient interval is a plurality of values for a corresponding plurality of cepstral coefficient intervals.
10 . The method as recited in claim 9 , wherein the plurality of cepstral coefficient intervals comprises a number of cepstral coefficient intervals in a range from about 2 to about 40 cepstral coefficient intervals, wherein, optionally,
the MDDSO is a click; and
the plurality of cepstral coefficient intervals spans a cepstral coefficient range from about 0 to about 12.
11 . The method as recited in claim 1 , wherein the certain parameter characterizing the sound in each time interval is a value output by a particular filter.
12 . The method as recited in claim 11 , wherein the value output by the particular filter is a plurality of values output by a corresponding plurality of filters.
13 . The method as recited in claim 12 , wherein the plurality of filters comprises about 10 to about 60 filters.
14 . The method as recited in claim 12 , wherein the plurality of filters comprises about 20 to about 40 filters.
15 . The method as recited in claim 1 , wherein the first duration is selected in a range from 10 milliseconds to 2000 milliseconds.
16 . The method as recited in claim 1 , wherein the first duration is selected in a range from 10 milliseconds to 2000 milliseconds and each time interval is selected in a range from 10 milliseconds to 130 milliseconds and the time interval is less than or equal to the first duration.
17 . The method as recited in claim 1 , wherein:
prior to receiving the operational data, the neural network has been trained further with third training data that indicates a plurality of time series of ambient sound having the first duration during operation of a different second type of medicament or test delivery device, wherein each time series of the third training data includes sound from a different second MDDSO emitted from the second type of medicament or test delivery device; and
the classification data further classifies the operational data as the first type of medicament or test delivery device or the second type of medicament or test delivery device or neither.
18 . The method as recited in claim 1 , further comprising, before receiving the operational data, performing the steps of:
receiving information indicating a type of medicament or test delivery device; and
retrieving configuration data that indicates configuration of the neural network for the type of medicament or test delivery device; and
configuring the processor based on the configuration data as the neural network for the type of medicament or test delivery device.
19 . A non-transitory computer-readable medium carrying one or more sequences of instructions, wherein execution of the one or more sequences of instructions by one or more processors configured as a neural network causes the one or more processors to perform the following steps:
receiving operational data that indicates a time series having a duration of at least a first duration for a signal from a vibration detector collected during operation of an instance of a type of medicament or test delivery device, with the type of medicament or test delivery device includes an operational component being operated by an operator undergoing training for administering a medicament;
generating classification data that classifies the operational data as a first medicament or test delivery device sound output (MDDSO) or not by inputting processed data based on the operational data into an input layer of the neural network; and
sending the classification data as a signal to an output device, with the classification data providing feedback to an operator of the type of medicament or test delivery device (MDD) to indicate whether the type of medicament or test delivery device was correctly operated when the operational data is classified as the first MDDSO, with the feedback being in terms of at least one of resistance pressure, measure of an amount of medicament delivered, and detect proper forces and sequence on operation of the operational component, and to display a pass or fail result;
wherein prior to receiving the operational data, the neural network has been trained based on the following:
collecting a first training time series corresponding to first training data that indicates a plurality of time series of ambient sounds, each first training time series having the first duration,
collecting a second training time series corresponding to second training data that indicates a plurality of time series of sound having the first duration during operation of the first type of medicament or test delivery device, each second training time series of the second training data includes sound from the first MDDSO emitted from the first type of medicament or test delivery device,
dividing the first duration of the first training data and the second training data into one or more window intervals,
processing each window interval into one or more values for a corresponding number of parameters characterizing sound,
processing acoustic time series data in each window interval to produce one or more values for a corresponding number of parameters characterizing sound in the window interval, and
providing the processed acoustic time series data in each window interval to an input layer of the neural network, and with an output layer of the neural network to produce correct classification data for the one or more values of the processed acoustic time series data provided to the input layer.