AI-based method for determining oxygen saturation level

Abstract

The specific implementations described herein disclose an artificial intelligence (AI)-based method for generating an oxygen saturation level output signal using a trained neural network. In one specific implementation, the method includes receiving a photoplethysmographic (PPG) signal, the PPG signal including a red PPG signal and an infrared PPG signal; generating an input feature matrix by performing a time-frequency transform of the PPG signal; training a neural network using the input feature matrix and an oxygen saturation level input signal; and generating an oxygen saturation level output signal using the trained neural network.

Description

Background Technology

[0001] Oximetry is an optical method for measuring oxyhemoglobin in the blood. It is based on the ability of different forms of hemoglobin to absorb light at different wavelengths. Oxyhemoglobin (HbO2) absorbs most strongly at red wavelengths, while deoxyhemoglobin or reduced hemoglobin (RHb) absorbs most strongly at near-infrared wavelengths. The transmittance of each wavelength as red and infrared light passes through blood vessels is inversely proportional to the concentrations of HbO2 and RHb in the blood. Pulsating oximeters can distinguish between alternating light inputs from arterial pulsations and the constant-level contributions from veins and other non-pulsating elements. Typically, only alternating light inputs are selected for analysis. Pulsating oximetry has proven to be a highly accurate technique. Pulsating oximeters typically provide arterial oxygen saturation, heart rate, and photoplethysmography (PPG) signals, such as red and infrared signals. Summary of the Invention

[0002] The specific implementation described herein discloses an artificial intelligence (AI)-based method for generating an oxygen saturation level output signal using a trained neural network. In one implementation, the method includes receiving a photoplethysmography (PPG) signal, which includes a red PPG signal and an infrared PPG signal; generating one or more input feature matrices by performing a time-frequency transformation of the PPG signal; training a neural network using the input feature matrices and an oxygen saturation level input signal; and using the trained neural network to generate an oxygen saturation level output signal.

[0003] A method for determining an oxygen saturation level includes receiving a photoplethysmography (PPG) signal, comprising a red PPG signal and an infrared PPG signal; generating an input feature matrix by performing a time-frequency transform of the PPG signal; training a neural network using the input feature matrix and an oxygen saturation level input signal; and using the trained neural network to generate an oxygen saturation level output signal. In one embodiment, generating the input feature matrix by performing a time-frequency transform of the PPG signal further includes generating the input feature matrix by performing a wavelet transform of the PPG signal.

[0004] In one specific implementation, generating the input feature matrix by performing a time-frequency transform of the PPG signal further includes generating modulus and phase vectors across the time-frequency plane. Alternatively, generating the input feature matrix by performing a time-frequency transform of the PPG signal further includes generating real and imaginary vectors across the time-frequency plane. Still alternatively, generating the input feature matrix by performing a wavelet transform of the PPG signal further includes generating the input feature matrix by performing a Morlet wavelet transform of the PPG signal.

[0005] In one implementation, the method further includes normalizing the PPG signal using a baseline to generate a normalized PPG signal, wherein generating the input feature matrix further includes generating the input feature matrix by performing a time-frequency transform of the normalized PPG signal. Alternatively, the method further includes non-linearly rescaling the input feature matrix before training the neural network using the input feature matrix. Even more alternatively, non-linearly rescaling the input feature matrix further includes rescaling the input feature matrix using logarithmic scaling. Alternatively, performing a time-frequency transform of the PPG signal further includes performing one of the following: performing a short-time Fourier transform (STFT) of the PPG signal, performing a Wigner-Ville transform of the PPG signal, and performing an S-transform of the PPG signal. Alternatively, the method further includes combining two or more vectors of the input feature matrix to generate a combined feature vector, and wherein training the neural network further includes training the neural network using the combined feature vector.

[0006] In a computing environment, a method at least partially executed on at least one processor includes receiving a photoplethysmography (PPG) signal, the PPG signal including a red PPG signal and an infrared PPG signal; generating an input feature matrix by performing a time-frequency transformation of the PPG signal; training a neural network using the input feature matrix and an oxygen saturation level input signal; and generating an oxygen saturation level output signal using the trained neural network.

[0007] A physical article comprising one or more tangible computer-readable storage media encoding computer-executable instructions for performing a computer process on a computer system to provide an automated connection to a collaborative event for a computing device, the computer process comprising: receiving a photoplethysmography (PPG) signal including a red PPG signal and an infrared PPG signal; generating an input feature matrix by performing a time-frequency transformation of the PPG signal; and training a neural network using the input feature matrix and an oxygen saturation level input signal.

[0008] This synopsis is provided to introduce, in a simplified form, a series of concepts further described in the detailed embodiments described below. This synopsis is not intended to identify the principal or essential features of the subject matter of the claims, nor is it intended to limit the scope of the claims.

[0009] This document also describes and presents other specific implementations. Attached Figure Description

[0010] A further understanding of the nature and advantages of the invention can be achieved by referring to the accompanying drawings described in the remainder of the specification.

[0011] Figure 1 An exemplary schematic diagram of an AI-based system is shown, which uses time-frequency preprocessing to determine input features to determine oxygen desaturation levels.

[0012] Figure 2 An exemplary schematic diagram 200 is shown for the wavelet transform process used to generate the input feature matrix.

[0013] Figure 3 An exemplary depiction of the wavelet transform modulus and phase diagram derived from PPG is shown.

[0014] Figure 4 An exemplary schematic diagram shows the modulus matrices of red and infrared PPG signals that are input into a deep learning model.

[0015] Figure 5 An exemplary implementation of a deep learning network is shown.

[0016] Figure 6 An exemplary output of the deep learning model compared to the true oxygen saturation level is shown.

[0017] Figure 7 An alternative exemplary schematic diagram is shown, illustrating another exemplary input to a deep learning model, where the modulus and phase of both red and infrared signals are input into the network.

[0018] Figure 8 An exemplary depiction of the real and imaginary parts of the transform generated by the time-frequency converter disclosed herein is shown.

[0019] Figure 9 An exemplary transformation modulus rescaled logarithmically is shown.

[0020] Figure 10 An exemplary operation of the AI-based method for determining oxygen saturation levels disclosed herein is shown.

[0021] Figure 11 An exemplary computing system that can be used to implement the described technology is shown. Detailed Implementation

[0022] Hypoxemia is a condition indicating that the oxygen level in a patient's arterial blood is below normal. Hypoxemia leads to a state of hypoxia or low oxygen levels, characterized by insufficient oxygen content in the patient's tissues. A pulse oximeter can be used to measure the oxygen content of arterial blood to indicate existing hypoxia and predict impending hypoxia. Pulse oximeters generate various output signals, including a time series of fluctuations recorded by the pulse oximeter, or red and infrared photoplethysmography (PPG) signals. Red and infrared PPG signals can be processed to generate oxygen saturation levels, also known as SpO2 values.

[0023] Using deep learning for time series analysis, such as determining SpO2 values ​​from red and infrared PPG signals, may simply involve feeding the raw PPGs into a network, such as a Long Short-Term Memory (LSTM) network or a Convolutional Neural Network (CNN). However, improved results can be obtained by preprocessing the data to provide input features that more readily capture information from the PPG signals. The AI-based method disclosed in this paper for determining oxygen saturation levels involves time-frequency processing of red and infrared PPG signals to generate an input feature matrix, which is then fed into a machine learning model, such as a CNN-based deep learning AI model.

[0024] Figure 1 An exemplary schematic diagram of an AI-based system 100 is shown, which is used to determine oxygen desaturation levels by using time-frequency preprocessing to determine input features. A pulse oximeter 104 can be used to measure a patient's oxygen saturation (SpO2) level. For example, the pulse oximeter 104 can be attached to a patient's thumb. The pulse oximeter 104 can be communicatively connected to a computing system 130. For example, the pulse oximeter 104 can be wirelessly connected to the computing system 130, and it can transmit a sequence 110 of input signals measured by the pulse oximeter 104 over a period of time. For example, such a sequence of input signals 110 can be transmitted per second. In one specific implementation, the input signal sequence 110 may include PPG signals, such as a red signal 110a, an infrared signal 110b, etc. The pulse oximeter 104 can also use the values ​​of the red signal 110a and the infrared signal 110b to generate a value for the oxygen saturation level (SpO2 level). Such SpO2 level sequences generated by pulse oximeter 104 are shown as oxygen saturation level sequence 120 (or SpO2 level sequence 120).

[0025] The computing system 130 may be a computing system including a microprocessor 132 and various other components implemented on memory 134. (The following is a further explanation...) Figure 10Examples of such computing systems 130 are disclosed herein. In the methods disclosed herein, memory 134 can be used to store an input signal sequence 110 generated by pulse oximeter 104, an SpO2 level sequence 120 generated by pulse oximeter 104, and one or more input feature matrices 142 generated based on the input signal sequence 110. The combination of input matrices 142 may also be referred to as a tensor. For example, time-frequency converter 140 can be used to generate input feature matrices 142 based on the input signal sequence 110.

[0026] In one implementation, the time-frequency transformer 140 can be a wavelet transformer, which allows the signal to be decomposed so that the frequency characteristics and location of specific features in the time series can be highlighted simultaneously. The properties of the wavelet transformer make it well-suited for analyzing signals where higher frequencies require more precise time resolution than lower frequencies (such as PPG signals). Furthermore, when analyzing higher frequencies, the time signal can be effectively amplified by employing a variable-width window, thereby providing higher time resolution when necessary. In one implementation, the wavelet transform of a continuous real-valued time signal x(t) with respect to the wavelet function ψ is defined as:

[0027]

[0028] Where t is time, a is the dilation parameter, b is the position parameter, ψ((tb) / a) is the analytical wavelet used in the convolution, and ψ*((tb) / a) is its complex conjugate, and x(t) is the signal under study, which in this application may be the PPG signal 110 obtained from the pulse oximeter 104. Examples of various wavelets available for use by the time-frequency transformer 140 may include the Monet wavelet, the Mexican hat wavelet, the Paul wavelet, etc. Alternatively, a discrete wavelet transform with a corresponding discrete wavelet may also be used. In other specific implementations, variants of the same wavelet, such as the Monet wavelet with different characteristic (or center) frequencies, may also be used.

[0029] In one exemplary implementation using a Monet wavelet at time-frequency transformer 140, the center frequency ω0 is set to 5.5. However, alternative center frequencies, such as values ​​less than 5.5 (e.g., 2.5 or 1.5), can also be used, as they are better suited for extracting time information. Alternatively, in other implementations, values ​​greater than 5.5 (e.g., 10, 15) can be used as center frequencies, which are better suited for extracting low-frequency metrics from the signal. In another implementation, time-frequency transformer 140 can run a continuous wavelet transform on the input signal 110 using Monet wavelets set to two or more individual wavelet center frequency values ​​(ω0), and input the modulus and phase from these additional matrices as additional inputs to the neural network training phase 144a.

[0030] The output of the time-frequency converter 140 can be stored as an input feature matrix 142. The input feature matrix 142 can be a complex-valued matrix in the time-frequency plane. An example of such a matrix could be a modulus value matrix spanning the time-frequency plane. Another example of such a matrix could be a phase value matrix spanning the time-frequency plane. Such modulus and phase values ​​spanning the time-frequency plane are described below. Figure 2 As shown in the image.

[0031] In an alternative implementation, the time-frequency converter 140 may normalize the PPG signal 110 using a baseline before calculating the wavelet transform. Alternatively, the transform value output from the time-frequency converter 140 may be rescaled non-linearly to enhance subtle features within the signal. For example, the modulus value output from the time-frequency converter 140 may be logarithmically scaled, and the scaled value may be used as part of the input feature matrix 142.

[0032] Specific implementations of the time-frequency transformer 140 may use alternative transforms, such as the short-time Fourier transform (STFT), Wigner-Weil transform, S-transform, etc. Alternatively, such alternative transforms can be computed and combined with wavelet transforms or with each other, and this combined transform can be input into the neural network training phase 144a. Similarly, different time-frequency transforms can be used as additional inputs to the neural network training phase 144a. For example, the phase and modulus matrices from the wavelet transform plus the phase and modulus matrices from the short-time Fourier transform can be used as inputs to the neural network training phase 144a. Alternatively, only a portion of such transforms, such as the real part, imaginary part, modulus, or phase, can be input into the neural network training phase 144a.

[0033] The input feature matrix 142 and the oxygen saturation level sequence 120 generated by the pulse oximeter 104 are used as inputs to train a new neural network at the neural network training phase 144a. For example, the neural network can be a deep learning network, such as a CNN.

[0034] Once the neural network is trained, as indicated by the trained neural network 144b, the input feature matrix 142 can be fed into the trained neural network 144b to generate a predicted oxygen saturation level sequence 150. A comparison of the predicted oxygen saturation level sequence 150 and the oxygen saturation level sequence 120 generated by the pulse oximeter 104 is depicted at 160 (and further described below). Figure 5 (as shown in the image).

[0035] In one specific implementation, the various vectors input to the feature matrix 142 can be combined in some way before being input into the deep learning neural network 144a. For example, the modulus value of the wavelet transform of the red signal can be divided by the modulus value of the wavelet transform of the infrared signal. Subsequently, this transform ratio can be used as input to the deep learning neural network 144a. Other operations besides division can also be used.

[0036] Figure 2 An exemplary schematic diagram 200 is shown for the wavelet transform process used to generate the input feature matrix. Specifically, schematic diagram 200 shows the transformation of the input PPG signal 202 at 204 to generate a modulus matrix 206 and a phase matrix 208 in the time-frequency plane.

[0037] Figure 3 An exemplary depiction 300 of the wavelet transform modulus map 304 and phase map 306 derived from the PPG signal 302 is shown. The matrices representing the modulus map 304 and phase map 306 can be input into a deep learning network, such as a CNN, to derive the oxygen saturation level.

[0038] Figure 4 An exemplary schematic diagram shows the modulus matrices 400 of the red and infrared PPG signals input into a deep learning model. Specifically, modulus matrix 402 is generated from the time-frequency transform of the red PPG signal, and modulus matrix 404 is generated from the time-frequency transform of the infrared PPG signal. Both matrices are input into the deep learning network.

[0039] Figure 5 An exemplary implementation of a deep learning network 500 is shown. The deep learning network 500 shown is a convolutional neural network (CNN). Here, it is a series of blocks 502 that can be repeated any number of times.

[0040] Figure 6 It shows the use of from Figure 4 Figure 600 provides an exemplary comparison of the input of the modulus matrices of the red and infrared signals described in the figure, and the output of a deep learning model compared to the true oxygen saturation level. Specifically, Figure 6The diagram illustrates successive desaturation events over a period of time. As shown in this paper, dashed line 602 represents the true oxygen saturation level, and solid line 604 represents the predicted oxygen saturation level derived from the deep learning model.

[0041] Figure 7 An alternative exemplary schematic diagram of another exemplary input to the deep learning model is shown, in which the modulus and phase of both red and infrared signals are input to the network. Specifically, modulus matrix 702 is generated from the time-frequency transform of the red PPG signal, and modulus matrix 706 is generated from the time-frequency transform of the infrared PPG signal. Furthermore, phase matrix 704 is generated from the time-frequency transform of the red PPG signal, and phase matrix 708 is generated from the time-frequency transform of the infrared PPG signal. Each of these four matrices is input to the deep learning network. In an alternative embodiment, modulus and phase matrices from multiple wavelet transforms can be used for both red and infrared signals. For example, an embodiment can use multiple Monet wavelet transforms for each signal, where the characteristic frequency of the Monet analysis wavelet is changed for each transform. For example, if three characteristic frequencies are used, three modulus matrices and three phase matrices can be generated per signal, thus a total of twelve matrices, which are used as input to train the network.

[0042] Figure 8 An exemplary depiction 800 of the real and imaginary parts of a transform generated as disclosed herein by the time-frequency converter is shown. Specifically, the real part 804 and the imaginary part 806 of the transform constitute the real and imaginary parts of a composite number, thereby forming the transform matrix generated by the time-frequency transform of the short PPG signal 802. The real part 804 and the imaginary part 806 of the transform can be input into the neural network training phase 144a.

[0043] Figure 9 A time-frequency plot 900 of an exemplary transform modulus rescaled logarithmically is shown. Specifically, 902 contains the original modulus transform matrix, and 904 contains a logarithmically scaled modulus generated by rescaling the transform values ​​of the original transform matrix. This type of scaling allows smaller features of the original modulus transform matrix to be less dominated by higher energy features in the signal. Scaling methods other than logarithmic scaling can also be used to scale transform values ​​non-linearly to enhance subtle features within the signal.

[0044] Figure 10 An exemplary operation 1000 of the AI-based method for determining oxygen saturation levels disclosed herein is illustrated. Operation 1002 acquires red and infrared PPG signals from a pulse oximeter. Operation 1004 calculates the wavelet transform of the input red and infrared PPG signals to generate a transform matrix WT. R and WT IR At operation 1006, the transformation matrix WT R and WT IRIt is fed into a deep learning neural network to train the network to generate the calculated oxygen saturation level, or SpO2.

[0045] Figure 11 An exemplary system 1100 is shown that can be used to implement the described techniques for providing provable and destructible device identity. (The system is intended for implementing the described techniques.) Figure 11 Exemplary hardware and operating environments include: computing devices (such as general-purpose computing devices in the form of a computer 20), mobile phones, personal data assistants (PDAs), tablets, smartwatches, game controllers, or other types of computing devices. Figure 11 In a specific implementation, for example, computer 20 includes a processing unit 21, system memory 22, and a system bus 23, which communicatively couples various system components, including the system memory, to the processing unit 21. There may be only one or more processing units 21, such that the processor of computer 20 includes a single central processing unit (CPU) or multiple processing units, commonly referred to as a parallel processing environment. Computer 20 may be a conventional computer, a distributed computer, or any other type of computer; however, the specific implementation is not limited to this.

[0046] System bus 23 can be any of several types of bus architectures, including memory bus or memory controller, peripheral bus, switching architecture, point-to-point connector, and local bus using any of the various bus architectures. System memory can also be simply referred to as memory and includes read-only memory (ROM) 24 and random access memory (RAM) 25. Basic input / output system (BIOS) 26, containing basic routines such as those that facilitate the transfer of information between components within computer 20 during startup, is stored in ROM 24. Computer 20 further includes hard disk drive 27 (not shown) for reading from and writing to a hard disk, disk drive 28 for reading from or writing to a removable disk 29, and optical disk drive 30 for reading from or writing to a removable optical disk 31, such as a CD-ROM, DVD, or other optical media.

[0047] Hard disk drive 27, disk drive 28, and optical disk drive 30 are connected to system bus 23 via hard disk drive interface 32, disk drive interface 33, and optical disk drive interface 34, respectively. The drives and their associated tangible computer-readable media provide non-volatile storage of computer-readable instructions, data structures, program modules, and other data of computer 20. Those skilled in the art will understand that any type of tangible computer-readable media can be used in the exemplary operating environment.

[0048] Multiple program modules may be stored on hard disk 27, disk 28, optical disk 30, ROM 24, or RAM 25, including an operating system 35, one or more applications 36, other program modules 37, and program data 38. Users can generate prompts on the personal computer 20 via input devices such as a keyboard 40 and pointing device 42. Other input devices (not shown) may include a microphone (e.g., for voice input), a camera (e.g., for a Natural User Interface (NUI)), a joystick, a gamepad, a dish satellite dish, a scanner, etc. These and other input devices are typically connected to the processing unit 21 via a serial port interface 46 coupled to the system bus 23, but may also be connected via other interfaces such as a parallel port, a game port, or a Universal Serial Bus (USB) (not shown). A monitor 47 or other type of display device is also connected to the system bus 23 via an interface such as a video adapter 48. In addition to the monitor, the computer typically includes other peripheral output devices (not shown) such as speakers and printers.

[0049] Computer 20 may operate in a networked environment using logical connections to one or more remote computers, such as remote computer 49. These logical connections are implemented through communication devices coupled to computer 20 or a portion thereof; the specific implementation is not limited to a particular type of communication device. Remote computer 49 may be another computer, server, router, network PC, client, peer device, or other common network node, and typically includes many or all of the elements described above with respect to computer 20. Figure 7 The logical connections described include local area networks (LANs) 51 and wide area networks (WANs) 52. Such networking environments are common in office networks, enterprise-wide computer networks, intranets, and the Internet, all of which are various types of networks.

[0050] When used in a LAN networking environment, computer 20 connects to local area network 51 via a network interface or adapter 53, which is a type of communication device. When used in a WAN networking environment, computer 20 typically includes a modem 54, a network adapter, a type of communication device for establishing communication over wide area network 52, or any other type of communication device. Modem 54 can be internal or external and is connected to system bus 23 via serial port interface 46. In a networking environment, the program engine relative to personal computer 20 or described in part therein may be stored in a remote memory storage device. It should be understood that the network connections shown are illustrative, and other means of communication devices for establishing communication links between computers may be used.

[0051] In an exemplary embodiment, software or firmware instructions for providing a provable and destructible device identity may be stored in memory 22 and / or storage device 29 or 31 and processed by processing unit 21. One or more data repositories disclosed herein may be stored as permanent data repositories in memory 22 and / or storage device 29 or 31. For example, an AI-based SpO2 determination module 1102 (shown within personal computer 20) may be implemented on computer 20 (alternatively, the AI-based SpO2 determination module 1102 may be implemented on a server or in a cloud environment). The AI-based SpO2 determination module 1102 may utilize one or more of the processing unit 21, memory 22, system bus 23, and other components of personal computer 20.

[0052] Compared to tangible computer-readable storage media, intangible computer-readable communication signals may contain computer-readable instructions, data structures, program modules, or other data residing in modulated data signals, such as carrier waves or other signal transmission mechanisms. The term "modulated data signal" means a signal having one or more of its characteristics set or altered in such a manner as encoding information in the signal. By way of example and not limitation, intangible communication signals include wired media such as wired networks or direct wired connections, and wireless media such as acoustic, RF, infrared, and other wireless media.

[0053] The specific implementations described herein are implemented as logical steps in one or more computer systems. Logical operations can be implemented as: (1) a sequence of processor-implemented steps executed in one or more computer systems, and (2) interconnected machines or circuit modules within one or more computer systems. Implementation is a matter of choice, depending on the performance requirements of the computer system utilized. Therefore, the logical operations constituting the specific implementations described herein are referred to differently as operations, steps, objects, or modules. Furthermore, it should be understood that logical operations can be performed in any order unless otherwise expressly stated or the language of the claims inherently requires a specific order.

[0054] The foregoing specification, examples, and data provide a complete description of the structure and use of exemplary embodiments of the present invention. Since many specific embodiments of the invention can be made without departing from the spirit and scope of the invention, the invention is contained in the appended claims. Furthermore, structural features of different embodiments can be combined in yet another embodiment without departing from the stated claims.

Claims

1. A method of determining an oxygen saturation level, the method comprising: receiving a photoplethysmographic (PPG) signal, the PPG signal comprising a red PPG signal and an infrared PPG signal; generating an input feature matrix by performing a time-frequency transform of the PPG signal, wherein generating the input feature matrix further comprises generating a real-valued vector and an imaginary-valued vector across a time-frequency plane; training a neural network using the input feature matrix, an oxygen saturation level input signal, and the real-valued vector and the imaginary-valued vector across the time-frequency plane; and generating an oxygen saturation level output signal using the trained neural network.

2. The method of claim 1, wherein generating the input feature matrix by performing the time-frequency transform of the PPG signal further comprises generating the input feature matrix by performing a wavelet transform of the PPG signal.

3. The method of claim 2, wherein generating the input feature matrix by performing the time-frequency transform of the PPG signal further comprises generating a magnitude-valued vector and a phase-valued vector across the time-frequency plane.

4. The method of claim 2, wherein generating the input feature matrix by performing the wavelet transform of the PPG signal further comprises generating the input feature matrix by performing a Morlet wavelet transform of the PPG signal.

5. The method of claim 1, further comprising normalizing the PPG signal by a baseline to generate a normalized PPG signal, wherein generating the input feature matrix further comprises generating the input feature matrix by performing a time-frequency transform of the normalized PPG signal.

6. The method of claim 1, further comprising nonlinearly rescaling the input feature matrix prior to training the neural network using the input feature matrix.

7. The method of claim 6, wherein nonlinearly rescaling the input feature matrix further comprises rescaling the input feature matrix using a logarithmic scaling.

8. The method of claim 1, wherein performing the time-frequency transform of the PPG signal further comprises one of performing a short-time Fourier transform (STFT) of the PPG signal, performing a Wigner-Ville transform of the PPG signal, and performing an S-transform of the PPG signal.

9. The method of claim 1, further comprising combining two or more vectors of the input feature matrix to generate a combined feature vector, and wherein training the neural network further comprises training the neural network with the combined feature vector.

10. In a computing environment, a method executed at least partially on at least one processor, the method comprising: receiving a photoplethysmographic (PPG) signal, the PPG signal comprising a red PPG signal and an infrared PPG signal; generating an input feature matrix by performing a time-frequency transform of the PPG signal, wherein generating the input feature matrix further comprises generating a real-valued vector and an imaginary-valued vector across a time-frequency plane; training a neural network using the input feature matrix, an oxygen saturation level input signal, and the real-valued vector and the imaginary-valued vector across the time-frequency plane; and generating an oxygen saturation level output signal using the trained neural network. using the trained neural network to generate an oxygen saturation level output signal.

11. The method of claim 10, wherein generating an input feature matrix by performing a time-frequency transform of the PPG signal further comprises generating an input feature matrix by performing a wavelet transform of the PPG signal.

12. The method of claim 11, wherein generating the input feature matrix by performing a time-frequency transform of the PPG signal further comprises generating a modulus value vector and a phase value vector across a time-frequency plane.

13. The method of claim 11, wherein generating the input feature matrix by performing a wavelet transform of the PPG signal further comprises generating the input feature matrix by performing a Morlet wavelet transform of the PPG signal.

14. The method of claim 10, further comprising rescaling the input feature matrix using a logarithmic scaling prior to training a neural network using the input feature matrix.

15. The method of claim 14, wherein performing a time-frequency transform of the PPG signal further comprises one of: performing a short-time Fourier transform (STFT) of the PPG signal, performing a Wigner-Ville transform of the PPG signal, and performing an S-transform of the PPG signal.

16. A physical article of manufacture comprising one or more tangible computer-readable storage media comprising computer-executable instructions embodied therein that, when executed by one or more computers, cause the computers to perform a computer process for providing a computing device with automated connectivity to a collaboration event, the computer process comprising: receiving a photoplethysmographic (PPG) signal, the PPG signal comprising a red PPG signal and an infrared PPG signal; generating an input feature matrix by performing a time-frequency transform of the PPG signal, wherein generating an input feature matrix further comprises generating a real value vector and an imaginary value vector across a time-frequency plane; and training a neural network using the input feature matrix, an oxygen saturation level input signal, and the real value vector and the imaginary value vector across the time-frequency plane.

17. The physical article of manufacture of claim 16, wherein the computer process further comprises using the trained neural network to generate an oxygen saturation level output signal.

18. The physical article of manufacture of claim 17, wherein generating an input feature matrix by performing a time-frequency transform of the PPG signal further comprises generating an input feature matrix by performing a wavelet transform of the PPG signal.

19. The physical article of manufacture of claim 18, wherein generating the input feature matrix by performing a time-frequency transform of the PPG signal further comprises generating a modulus value vector and a phase value vector across a time-frequency plane.

Images