Method and apparatus for conversion between random signals

Abstract

The present disclosure relates to methods and devices for converting between random signals. A source random signal is decomposed into its intrinsic components using a decomposition process. The intrinsic components are transformed, and a set of machine learning models are defined and trained to process the individual components in the transformed components. Thus, the source random signal is empirically decomposed into basis components, which are then used as a learning data set for the set of machine learning models to predict target components. The target components are then individually predicted and combined to reconstruct the predicted target random signal. The source random signal and the target random signal can be biological signals having a related or common origin, such as photoplethysmogram signals and arterial blood pressure waveforms.

Description

Technical Field

[0001] The present invention relates to the conversion of a first random signal into a second random signal. In particular, but not exclusively, the present invention relates to the automatic conversion of a first random biological signal into a second random biological signal; more specifically, wherein the first random biological signal is a photoplethysmogram or an electrocardiogram signal, and the second random biological signal is an arterial blood pressure waveform. Background Art

[0002] In current clinical practice, blood pressure measurement is either invasive using an arterial catheter or non-invasive using a cuff by oscillometry or auscultation. Invasive measurement is inherently continuous but risky (e.g., infection, bleeding, thrombosis). Therefore, invasive measurement is only used for critically ill patients and not for most patients with chronic hypertension. Intermittent cuff inflation is the gold standard for blood pressure measurement in patients with chronic hypertension. It is non-invasive but does not allow continuous blood pressure measurement.

[0003] Photoplethysmogram (PPG) is a non-invasive circulatory signal related to the pulsatile volume of blood in tissue and is typically acquired by a pulse oximeter. Blood pressure can be estimated from the PPG signal, which can be obtained from various devices. Such devices include mobile or wearable devices such as smartwatches, activity trackers, etc. However, the PPG signal is prone to artifacts that can negatively impact measurement accuracy. These artifacts can lead to a large number of misleading diagnoses. Due to increased movement of mobile devices, these artifacts are more pronounced in PPG signals collected by mobile devices. The accuracy of readings can vary depending on factors such as blood flow, device placement, and the stillness of the device wearer.

[0004] Blood pressure estimation is typically treated as a regression problem, i.e., extracting features from a combination of biological signals, which usually includes PPG and electrocardiogram (ECG). Features can also include signals obtained from activity or environmental sensors such as accelerometers, pressure sensors, oxygen sensors, etc. These features are then input into a machine learning algorithm to predict blood pressure. Machine learning algorithms applicable to blood pressure prediction include linear regression, support vector regression, Bayesian regression, and regression-based deep neural networks. In the regression process, these techniques can predict systolic blood pressure (SBP), diastolic blood pressure (DBP), and / or mean arterial pressure (MAP). Classification methods such as logistic regression can also be used to predict classification metrics such as normal, prehypertension, stage 1 hypertension, and stage 2 hypertension.

[0005] One problem with the above method is that fine features, such as double-spin notch and double-spin peak, do not always exist in the PPG signal and are difficult to capture reliably. The extraction of these features is also computationally expensive. Another problem is that motion-corrupted signals are difficult to repair. Even if fine features such as double-cycle notch and double-cycle peak are extracted, motion corruption will increase errors. Therefore, most existing estimation techniques directly disqualify motion-corrupted signals. Summary of the Invention

[0006] In the present invention, a decomposition process is used to deconstruct a source random signal into its intrinsic components. The intrinsic components are transformed, and a set of machine learning models is defined and trained to process each component in the transformed components. Thus, the source random signal is empirically decomposed into base components, which are then used as a learning dataset for the set of machine learning models to predict target components. The target components are then predicted and combined individually to reconstruct the predicted target random signal. The source random signal and the target random signal can be biological signals with related or common origins, such as photoplethysmogram signals and arterial blood pressure waveforms.

[0007] The present invention provides a method and apparatus for translating random signals, particularly biological signals with related origins. A source signal is deconstructed into a plurality of deconstructed source components, and a transformation is performed on the plurality of deconstructed source components to generate a corresponding plurality of transformed source components. A plurality of machine learning models are used to map the plurality of transformed source components to a plurality of estimated target components. Each of the plurality of machine learning models is trained to map a corresponding one of the plurality of transformed source components to a corresponding one of the plurality of estimated target components. An inverse transformation is performed on the plurality of estimated target components to generate a corresponding plurality of transformed target components. A target signal is generated by combining the plurality of transformed target components.

[0008] Thus, aspects of the present invention have prior information about the composition of the input and output signals, which helps to eliminate any data-dependent bias by enabling the machine learning models to learn the basic components of the signal rather than the entire signal. The underlying components are less complex than the original target signal, so the machine learning models are easier to learn the transfer function. This allows the use of a simpler and / or smaller model than predicting the original target signal. Using a simpler and / or smaller model may help to more effectively utilize storage space and consume less electrical energy during use, both of which are particularly advantageous for the use of low-power devices such as wearable devices, activity trackers, or smart watches.

[0009] Further features and aspects of the present invention are provided in the appended claims. Brief Description of the Drawings

[0010] The present invention will now be described by way of example only with reference to the accompanying drawings, in which:

[0011] Figure 1 Illustrates a process for converting a source signal into a target signal according to an aspect of the present invention;

[0012] Figure 2 Illustrates a wavelet-based multi-resolution analysis for signal denoising according to an embodiment of the present invention;

[0013] Figure 3 Illustrates an example empirical mode decomposition according to an embodiment of the present invention;

[0014] Figure 4 Illustrates an example U-net architecture suitable for use with embodiments of the present invention;

[0015] Figure 5 Illustrates a formula for summing transformed target components to generate a target signal according to an embodiment of the present invention;

[0016] Figure 6 Illustrates a method for converting a source signal into a target signal according to an aspect of the present invention;

[0017] Figure 7 Illustrates a process for converting a source signal into multiple components according to an embodiment of the present invention;

[0018] Figure 8 Illustrates a process for converting multiple estimated components into a target signal according to an embodiment of the present invention;

[0019] Figure 9 Illustrates a device suitable for performing embodiments of the present invention; and

[0020] Figure 10 Illustrates a wearable device suitable for performing embodiments of the present invention. Detailed Description

[0021] Many applications in the field of signal processing involve random signals that have a related or common origin. For example, a first cardiac signal and a second different cardiac signal both have a common origin in the contraction and dilation of various heart chambers and the blood pumping action. The present invention relates to the automatic conversion of random signals that have a related or common origin.

[0022] The present invention uses a decomposition process such as empirical mode decomposition (EMD) to decompose a source random signal into its intrinsic components, such as the intrinsic mode functions (IMFs) of the signal. For example, the decomposed components are converted to a suitable time-frequency representation, and a set of machine learning models are defined and trained to operate on individual components rather than the entire source random signal (as is typical in prior art methods). Although traditional machine learning models only accept one type of input signal, one advantage of the present invention is that each segment of the source random signal is decomposed into basic components empirically and then used as a learning data set for the set of machine learning models to predict the target component. The target component is predicted separately and then combined to reconstruct the predicted target signal.

[0023] One aspect of the present invention applies the above process to the conversion of biological random signals to achieve cuffless blood pressure prediction. By combining signal processing and machine learning, non-invasive photoplethysmogram (PPG) or electrocardiogram (ECG) signals are converted into a continuous arterial blood pressure (ABP) waveform. Such an ABP waveform can be used to further estimate systolic blood pressure, diastolic blood pressure, and central blood pressure values.

[0024] Thus, the present invention addresses the need for home, continuous, and non-invasive blood pressure monitoring that is reliable across population differences. While some prior art uses PPG and ECG signals to address this problem, the present invention illustrates physiological relevant parameters, separately extracts slow-changing and highly dynamic random signal components to predict the entire ABP waveform, and only requires PPG measurements for prediction.

[0025] Figure 1 FIG. 100 shows a process 100 for converting a source signal 102 into a target signal 104 according to one aspect of the present invention. Generally, process 100 includes a deconstruction stage 106, a mapping stage 108, and a reconstruction stage 110.

[0026] As will be described in more detail below, multiple components 112, 114 are generated from source signal 102 and subsequently mapped to corresponding multiple estimated components 116, 118 using multiple trained machine learning models 120, 122. Then, target signal 104 is generated from the multiple estimated components 116, 118.

[0027] Advantageously, by decomposing the source signal into individual intrinsic components and training a set of machine learning models on each component, the present invention can help ensure that the "learning" of each machine learning model has significant nuances (random allowable variations) at the level of time and frequency characteristics. This can be achieved by separately capturing and modeling highly dynamic random signal components rather than capturing and modeling the entire signal to more accurately estimate the target signal.

[0028] In one example, both the source signal 102 and the target signal 104 are random signals with a related or common origin. Although the present invention is applicable to the translation of any suitable random signal with a related origin, the present invention is mainly directed to biological signals with a related or common origin, and more specifically, relates to biological signals for cuffless blood pressure prediction. In such a setting, the source random signal is a PPG or ECG signal, and the target random signal is an ABP signal. In this case, the PPG or ECG signal and the ABP signal have a common origin in the contraction and dilation of various heart chambers and the blood pumping action. Thus, in the following description, in some cases, a reference to the source signal may refer to a PPG or ECG signal, while a reference to the target signal may refer to an ABP signal.

[0029] The source signal 102 can be obtained from an input signal (not shown). The input signal can be a continuous signal or waveform obtained from a sensor such as a PPG sensor or an ECG sensor. In one embodiment, the source signal 102 includes an interval of a predefined length obtained from the input signal. The predefined length defines the time length of the source signal 102, such as 2 seconds, 4 seconds, 8 seconds, etc. Thus, although the source signal 102 can represent a discrete signal corresponding to an interval of the predefined length, the processes and functions described herein can be repeatedly executed on continuous intervals obtained from the input signal to reproduce a continuous target signal.

[0030] To maintain important signal features while removing noise, the input signal can be preprocessed before obtaining the source signal 102. Alternatively, it can be preprocessed after obtaining the source signal 102. In one example, the preprocessing of the signal at least includes denoising the signal. Thus, the input signal can be denoised before obtaining the source signal 102.

[0031] In one example, the denoising is based on the wavelet decomposition of the signal (i.e., the wavelet decomposition of the input signal or the source signal 102). Advantageously, wavelet decomposition of the signal helps to improve the phase response. This can provide improved efficiency in terms of computational complexity, adaptability in different signal-to-noise ratio regions, and non-stationary artifacts. Wavelet-based multi-resolution analysis can decompose a time series such as the input signal or the source signal 102 into a set of component sequences with a well-defined hierarchical structure, as Figure 2 shown.

[0032] Figure 2A signal 202 (which can be an input signal or a source signal 102) is shown decomposed into a hierarchical structure 204, which includes approximation coefficients 204-1 and detail coefficients 204-2 across multiple levels. The decomposition is performed according to scale and direction. A wavelet-based multi-resolution analysis can be implemented using a cascaded filter bank 206, where a low-pass filter (e.g., low-pass filter 208) and a high-pass filter (e.g., high-pass filter 210) satisfy certain specific constraints. Due to the multi-scale decomposition, important signal features can be retained while removing noise.

[0033] Referring again to Figure 1 , after optionally preprocessing the input signal or source signal 102, a plurality of components 112, 114 are generated from the source signal 102 in the deconstruction stage 106. The deconstruction stage 106 corresponds to a transformation process of generating a plurality of components 112, 114 from the source signal 102. Therefore, the plurality of components 112, 114 can alternatively be referred to as a plurality of transformed components. Generally, the transformation process performed during the deconstruction stage 106 can correspond to any suitable transformation process or transformation function that has a corresponding inverse transformation process or function that can be performed during the reconstruction stage 110.

[0034] Therefore, the present invention empirically decomposes the source signal 102 into base components, namely a plurality of components 112, 114. These underlying source components are mapped to corresponding underlying target components, which can be reconstructed to form a target signal. By deconstructing and mapping the individual components of the signal rather than mapping the original source signal, finer characteristics and features of the source signal can be modeled and subsequently reconstructed in the target signal. Therefore, the present invention can provide a more accurate and robust process for transforming between random signals.

[0035] In one aspect, the transformation process performed during the deconstruction stage 106 corresponds to deconstruction 124 and transformation 126. Deconstruction 124 deconstructs the source signal 102 into a plurality of deconstructed source components 128, 130, and transformation 126 of the plurality of deconstructed source components 128, 130 generates a corresponding plurality of transformed source components corresponding to the plurality of components 112, 114. That is, the first component 112 and the second component 114 among the plurality of components 112, 114 can be generated by first decomposing or deconstructing the source signal 102 into a first deconstructed source component 128 and a second deconstructed source component 130. The first component 112 can be generated from the first deconstructed source component 128 using a transformation function, and the second component 114 can be generated from the second deconstructed source component 130 using a transformation function. Therefore, transformation 126 can correspond to a specific transformation function applied to the decomposed or deconstructed source components generated during deconstruction 124.

[0036] Optionally, the transformation process further includes generating a third component from the source signal. In one example, the third component is generated from the source random signal through the above-described steps of decomposition 124 and transformation 126 for the first component 112 and the second component 114. That is, the decomposition 124 can produce a third decomposed source component, and the third component is produced by the transformation 126 of the third decomposed source component.

[0037] In a particular aspect, the decomposition 124 of the source signal 102 into multiple decomposed source components 128, 130 is performed using empirical mode decomposition (EMD), and the transformation 126 of the multiple decomposed source components 128, 130 is performed using the Hilbert transform.

[0038] Empirical mode decomposition decomposes the source signal into a set of fully data-adaptive decomposed components. The decomposed components can also be referred to as components, decomposition components, basis functions, or intrinsic mode functions (IMFs). The basic theory of EMD is that any data set is essentially composed of a finite number of different simple oscillation modes (i.e., components or IMFs). Advantageously, compared with methods such as Fourier transform and wavelet-based transform, EMD has no prior-defined basis. Therefore, EMD can effectively process non-linear and non-stationary data. This can be particularly useful for random biological signals such as PPG or ECG signals.

[0039] Figure 3 An example of the empirical mode decomposition of the signal 302 is shown. The signal 302 can be considered equivalent to Figure 1 the source signal 102 shown, and can be a random signal, such as a PPG signal or an ECG signal.

[0040] In Figure 3 it, the signal 302 has been decomposed into a first decomposition or decomposed component 304, a second decomposition or decomposed component 306, a third decomposition or decomposed component 308, and a fourth decomposition or decomposed component 310 using EMD. The residue 312 is also shown. In the present invention, the components generated from EMD are considered to include the residue.

[0041] Referring again to Figure 1, during the decomposition stage 106, multiple transformations 126 of the decomposition source components 128, 130 can be performed to generate corresponding multiple transformed source components, which can correspond to the multiple components 112, 114 output from the decomposition stage 106. In an embodiment where EMD is used for the decomposition 124 of the source signal 102 and then for the transformation 126 of the multiple decomposition source components 128, 130, it can include the Hilbert transform of each decomposition source component (i.e., performing the Hilbert transform on the first decomposition source component 128 and independently performing the Hilbert transform on the second decomposition source component 130). As is well known, the Hilbert transform forms an analytic signal into a complex time signal corresponding to the spectral content of the positive frequency axis. Therefore, the Hilbert transform allows the calculation of the instantaneous properties of the decomposition source components, especially the envelope amplitude and the instantaneous frequency.

[0042] In an alternative embodiment, the transformation process performed during the decomposition stage 106 corresponds to the Hilbert-Huang transform (HHT) of the source signal 102, thereby generating the multiple components 112, 114 by performing HHT on the source signal 102.

[0043] HHT has been applied to many scientific research, such as biomedical signal processing, geophysics, image processing, structural testing, fault diagnosis, nuclear physics, etc. Like other spectral techniques, HHT is able to represent the amplitude and the instantaneous frequency in a three-dimensional plot or image, where the amplitude is the height on the time-frequency plane.

[0044] Therefore, when using the Hilbert transform or HHT, the transformed source components, such as the first component 112 and the second component 114, are the energy-frequency-time distributions that can be represented as three-dimensional vectors or images. Therefore, using the Hilbert transform or HHT can represent the components with high-resolution time and frequency, which can help improve the accuracy of the signal being reconstructed.

[0045] After the multiple components 112, 114 (i.e., multiple transformed source components) are generated in the decomposition stage 106, the multiple components 112, 114 are mapped to the corresponding multiple estimated components 116, 118 in the mapping stage 108. The multiple estimated components 116, 188 can alternatively be referred to as multiple estimated target components.

[0046] Prediction models, such as the first trained machine learning model 120 and the second trained machine learning model 122, are used to predict a set of estimated target components, such as the multiple estimated components 116, 118, from a set of transformed source components (such as components 112, 114). In one example implementation, the prediction model further includes a third trained machine learning model 123 for predicting a third estimated component from a third component.

[0047] A prediction model is trained to map components of a source signal to corresponding components of a target signal. That is, a single prediction model, such as a single machine learning model or neural network, can be trained using a dataset of known source signals and corresponding target signals (e.g., a dataset of PPG / ECG signals with known corresponding ABP waveforms). The PPG / ECG signals and corresponding ABP waveforms can be deconstructed into individual components so that the prediction model can be trained to predict from individual PPG / ECG components to corresponding ABP components (e.g., trained to predict the first component or IMF of the ABP waveform from the first component or IMF of the PPG waveform). As previously described, training a set of machine learning models on a per-component basis may help ensure that the "learning" of individual machine learning models is highly nuanced in terms of temporal and frequency characteristics. This can more accurately estimate the target signal by separately capturing and modeling highly dynamic random signal components rather than capturing and modeling the original source and target signals. Additionally, the base components are more complex than the original signals, so the prediction model can more easily learn the transfer function (i.e., learn to map from source components to estimated target components). This allows for the use of a simpler and / or smaller model than predicting the original target signal from the original source signal. This can be particularly beneficial when the present invention is embodied in a low-energy device such as a wearable device.

[0048] As will be described in more detail below, the prediction model used during mapping phase 108 can correspond to any suitable machine learning model trained to predict an estimated target component from a source component. In one example implementation, the first trained machine learning model 120 and the second trained machine learning model 122 are trained encoder-decoder neural networks, such as a trained U-net model or a trained VGG model. Other suitable neural network architectures include language translation networks, recurrent neural networks, convolutional neural networks, etc. Thus, the prediction model can be any suitable trained machine learning model or trained deep learning model.

[0049] A single machine learning model is trained to predict a corresponding estimated target component from a corresponding source component. For example, if the source signal is converted into four source components, the first machine learning model is trained to map the first source component to the first estimated target component, the second machine learning model is trained to map the second source component to the second estimated target component, and so on. Optionally, predicting the estimated target component can involve multiple source components provided to the machine learning model. In this case, mapping the first component to the first estimated component includes providing the first component and the second component as inputs to the first trained machine learning model. Additionally, or alternatively, mapping the second source component to the second estimated target component includes providing the first source component and the second source component as inputs to the second trained machine learning model.

[0050] A trained machine learning model can be trained using any suitable dataset specific to the underlying task and a training method specific to the underlying machine learning algorithm used. In one example implementation described below, the machine learning model includes a U-net model trained on the MIMIC-III dataset.

[0051] Figure 4 An example U-net architecture 400 applicable to the present invention is shown. The example U-net architecture 400 is based on the VGG-16 architecture and uses a dedicated decoder network rather than the minimal decoder used in a fully convolutional network (FCN). U-nets utilize unpooling layers to recover spatial information and make extensive use of skip connections. These connections utilize the spatial information in the encoding path and the concatenation of feature maps.

[0052] The MIMIC-III database is a large free database containing uncurated health-related data associated with more than 40,000 patients admitted to the intensive care unit at Beth Israel Deaconess Medical Center during the period from 2001 to 2021. The U-net model is trained on a subset of the MIMIC-III waveform database that contains 22,317 waveform records and 22,247 numeric records from 10,282 different ICU patients. These records typically include digital signals such as ECG, ABP, respiration, and PPG. These records also contain measurements such as heart rate, blood oxygen saturation, systolic blood pressure, mean pressure, and diastolic blood pressure. This database is a subset of the MIMIC-III waveform database representing records of identified patients whose corresponding clinical records can be found in the MIMIC-III clinical database. The training data obtained from the subset of the MIMIC-III waveform database match includes training instances corresponding to PPG signals and training targets corresponding to ABP signals. Both the training instances and the training targets contain 14,896 waveforms of length 1024. In the currently described implementation, using a suitable transformation process such as the Hilbert-Huang transform, the training instances and the training targets are decomposed into four transformed basis functions and accompanying residuals. In an alternative implementation, the number of transformed basis functions can be greater than or less than four, and other suitable transformation processes can be used. Five U-net neural networks are trained to map between five underlying signal components, i.e., between the training instances (PPG components or basis functions) and the training targets (ABP components or basis functions). For example, the first U-net neural network is trained to map the first PPG basis function to the first ABP basis function. Those skilled in the art will understand that the number of neural networks, or more generally the machine learning models, corresponds to the number of decomposed components. Thus, in an alternative implementation, the number of neural networks or machine learning models can be greater than or less than 5. In the currently described implementation, Figure 4The standard U-net architecture shown. Those skilled in the art will understand that alternative U-net architectures other than those shown can also be used. Figure 4 The U-net neural network is trained using the ADAM optimizer with mean absolute error loss. The training process is executed over 100 epochs with a batch size of 255 (with shuffling), and 30% of the training data is used for validation.

[0053] Referring again to Figure 1 , after multiple estimated components 116, 118 are predicted in the mapping stage 108, the target signal 104 is generated in the reconstruction stage 110.

[0054] Generally, in the reconstruction stage 110, a second transformation process or function corresponding to the inverse of the transformation process or function used during the deconstruction stage 106 is applied to generate the target signal 104 from the multiple estimated components 116, 118. In an implementation using the third machine learning model 123, the target signal 104 is generated from the first estimated component 116, the second estimated component 118, and the third estimated component.

[0055] In an example implementation, the inverse transformation process performed during the reconstruction stage 110 corresponds to transformation 132 and reconstruction 134. Transformation 132 transforms the multiple estimated components 116, 118 into corresponding multiple transformed target components 136, 138. Reconstruction 134 of the multiple transformed target components 136, 138 is performed to generate the target signal 104. A transformation function can be used to generate the first transformed target component 136 from the first estimated component 116, and a transformation function can be used to generate the second transformed target component 138 from the second estimated component 118. Thus, transformation 132 can correspond to a specific transformation function applied to the estimated target components. In one example, the transformation function applied at transformation 132 is the inverse function of the transformation function applied at transformation 126. In an implementation where a Hilbert transform is performed at transformation 126, then an inverse Hilbert transform is performed at transformation 132.

[0056] Reconstruction 134 of the multiple transformed target components 136, 138 combines the multiple transformed target components 136, 138 to generate the target signal 104. Any suitable method for combining the multiple transformed target components 136, 138 to generate the target signal 104 can be used. In one example implementation, the target signal 104 can be reconstructed by summing the multiple transformed target components 136, 138 (which preferably includes the corresponding residuals). Figure 5 An example summation according to this implementation is shown. The target signal 502 is determined to be the sum of multiple transformed target components including the residual 504 and a single transformed target component 506.

[0057] Optionally, the target signal 502 can be output for the user to view or further analyze. For example, the target signal 502 can be output to the display of a device or a wearable device. Alternatively, the target signal 502 can be sent to an external device for storage or further analysis / processing.

[0058] Advantageously, the present invention uses a trained machine learning model to predict the components of the decomposed target signal from the corresponding components of the decomposed source signal. Both the source signal and the target signal can be random signals with a related or common origin. The present invention can help improve the accuracy and quality of the predicted target random signal. In addition, since specific noise components are associated with the components of the decomposed source random signal, the present invention can provide improved noise processing. That is, decomposing the source random signal into a set of components can provide an inherent noise removal advantage for probabilistic noise. This may help eliminate the noise that the machine learning model does not need to learn, i.e., the potential noise from a common source, rather than just the noise generated by the measurement results. In addition, since the decomposed signal is much simpler than the source random signal, the present invention can provide improved phase prediction with respect to the phase of the "true" target random signal.

[0059] Figure 6 A method 600 for converting a source signal into a target signal according to an aspect of the present invention is shown. The method 600 includes steps 602, 604, 606, and optionally includes steps 601, 607.

[0060] Optionally, the method 600 first obtains 601 the source signal, such as Figure 1 the source signal 102 shown. In one example, obtaining 601 the source signal includes obtaining a source photoplethysmogram (PPG) signal. As described above, the source signal can be obtained from an input signal, which can be a continuous signal or a waveform obtained from a sensor (such as a PPG sensor or an ECG sensor). In one example, the source signal includes an interval of a predefined length obtained from the input signal. To maintain important signal features while removing noise, the input signal can be preprocessed before obtaining 601. Alternatively, the source signal can be preprocessed after obtaining 601. The preprocessing of the signal can include at least denoising the signal.

[0061] The method 600 includes generating 602 a plurality of components from the source signal. In one example, generating 602 a plurality of components from the source signal includes transforming the source signal into a plurality of components using a first transformation process. In another example, generating 602 a plurality of components from the source signal includes transforming the source PPG signal into a plurality of PPG components using a first transformation process.

[0062] Figure 7 A transformation process 700 that can be performed as part of generating 602 according to one example implementation is shown.

[0063] The transformation process 700 includes decomposing a source signal into a plurality of decomposed source components 702. In one example, the decomposition 702 includes performing empirical mode decomposition (EMD) of the source signal to generate a plurality of decomposed source signals.

[0064] The transformation process 700 also includes performing 704 a first transformation on each of the plurality of decomposed source components to generate a corresponding plurality of transformed source components. In an implementation where EMD is used for the decomposition 702, performing 704 the first transformation to transform the plurality of decomposed source components includes using the Hilbert transform as the first transformation. Thus, the Hilbert transform is applied to each decomposed source component among the plurality of decomposed target components.

[0065] Referring again to Figure 6 , in an alternative example, the first transformation process performed in the generation 602 step corresponds to the Hilbert-Huang transform. Alternatively, any suitable transformation process can be used to decompose the source signal into a plurality of components. The transformation process for decomposing the source signal can have a corresponding inverse transformation process that can be used to recover or reconstruct the target signal.

[0066] The method 600 also includes mapping 604 the plurality of components to a corresponding plurality of estimated components using a plurality of trained machine learning models. Each of the plurality of trained machine learning models can be trained to map a corresponding one of the plurality of components to a corresponding one of the plurality of estimated components. Thus, the mapping 604 can include predicting a first estimated component of a first component from the plurality of components using a first trained machine learning model, and predicting a second estimated component of a second component from the plurality of components using a second trained machine learning model.

[0067] As described in more detail above, the machine learning model used can correspond to any appropriately trained machine learning model, such as a trained U-net model or a trained VGG model. Other suitable neural network architectures include language translation networks, recurrent neural networks, convolutional neural networks, etc. An example training process for a U-net model applicable to the present invention was described above in connection with Figure 4 an example training process for a U-net model applicable to the present invention was described above in connection with

[0068] In one example implementation, the mapping 604 includes using a plurality of prediction models to predict a plurality of predicted arterial blood pressure (ABP) components from a plurality of PPG components. Each of the plurality of prediction models can be used to map from one or more of the plurality of PPG components to a predicted ABP component among the plurality of predicted ABP components.

[0069] Method 600 further includes generating 606 a target signal from a plurality of estimated components. In one example, generating 606 the target signal from the plurality of estimated components includes transforming the plurality of estimated components into the target signal using a second transformation process. In another example, generating 606 the target signal from the plurality of estimated components includes transforming the plurality of predicted ABP components into the target ABP signal.

[0070] Figure 8 Shows a conversion process 800 that can be performed as part of generating 606 according to one particular implementation.

[0071] The conversion process 800 includes performing 802 a second conversion process on each of the plurality of estimated target components to generate a plurality of transformed target components. The second transformation process can be the inverse of the first transformation process performed during source signal deconstruction (e.g., the first transformation process performed in step 704 of Figure 7 . In implementations where the Hilbert transform is used as the first transformation process, the inverse Hilbert transform can be used as the second transformation process.

[0072] The conversion process 800 further includes combining 804 the plurality of transformed target components to generate the target signal. As shown in Figure 5 and described in more detail above, combining 804 the plurality of transformed target components can correspond to the sum of the estimated target components (including residuals) to generate the target signal.

[0073] Referring again to Figure 6 , in an alternative implementation, when the first transformation process performed in generation step 602 corresponds to the Hilbert-Huang transform, the second transformation process performed in generation step 606 corresponds to the Hilbert-Huang transform.

[0074] Optionally, method 600 includes causing 607 the output of the target signal. In one example, causing 607 the output of the target signal includes causing the output of the target ABP signal. Additionally or alternatively, causing 607 the output of the target signal can include causing the target signal to be output to a display of a device or wearable device. Or, causing 607 the output of the target signal can include causing the target signal to be sent to an external device, e.g., for further processing.

[0075] Figure 9 shows a device suitable for performing the above aspects of the present disclosure. Figure 9 Shows a block diagram of an example computing system.

[0076] The computing system 900 can be configured to perform any of the operations disclosed herein. The computing system includes one or more computing devices 902. One or more computing devices 902 of the computing system 900 include one or more processors 904 and a memory 906. One or more processors 904 can be any general-purpose processor configured to execute a set of instructions (i.e., a set of executable instructions). For example, one or more processors 904 can be one or more general-purpose processors, one or more field-programmable gate arrays (FPGAs), and / or one or more application-specific integrated circuits (ASICs). In one example, one or more processors 904 include one processor. Alternatively, one or more processors 904 include multiple processors operatively connected. One or more processors 904 are communicatively coupled to the memory 906 via an address bus 908, a control bus 910, and a data bus 912. The memory 906 can be random access memory (RAM), read-only memory (ROM), a persistent storage device such as a hard disk drive, erasable programmable read-only memory (EPROM), etc. One or more computing devices 902 also include an input / output interface 914 communicatively coupled to the address bus 908, the control bus 910, and the data bus 912.

[0077] The memory 906 can store information accessible by one or more processors 904. For example, the memory 906 (e.g., one or more non-transitory computer-readable storage media, memory devices) can include computer-readable instructions (not shown) executable by one or more processors 904. The computer-readable instructions can be software written in any suitable programming language or can be implemented in hardware. Additionally, or alternatively, the computer-readable instructions can be executed on one or more processors 904 in logically and / or physically independent threads. For example, the memory 906 can store instructions (not shown) that, when executed by one or more processors 904, cause one or more processors 904 to perform operations such as any of the operations and functions for which the computing system 900 is configured, as described herein. Additionally, or alternatively, the memory 906 can store data (not shown) that can be obtained, received, accessed, written, manipulated, created, and / or stored. For example, the data can include data and / or information as described herein with respect to Figures 1 to 8 the description. In some implementations, the computing device 902 can obtain and / or store data from one or more memory devices remote from the computing system 900.

[0078] The computing system 900 also includes a storage unit 916, a network interface 918, an input controller 920, and an output controller 922. The storage unit 916, the network interface 918, the input controller 920, and the output controller 922 are communicatively coupled to the computing device 902 via the I / O interface 914.

[0079] The storage unit 916 is a computer-readable medium, preferably a non-transitory computer-readable medium or a non-transitory machine-readable medium, including one or more programs, and the one or more programs include instructions that, when executed by one or more processors 904, cause the computing system 900 to perform the method steps of the present invention. Alternatively, the storage unit 916 is a temporary computer-readable medium. The storage unit 916 can be a persistent storage device, such as a hard disk drive, a cloud storage device, or any other suitable storage device.

[0080] The network interface 918 can be a Wi-Fi module, a network interface card, a Bluetooth module, and / or any other suitable wired or wireless communication device. In one example, the network interface 918 is configured to connect to a network such as a local area network (LAN) or a wide area network (WAN), the Internet, or an intranet.

[0081] Figure 10 A wearable device suitable for performing the above aspects of the present invention is shown. In one example, the wearable device is a device for cuffless blood pressure estimation.

[0082] Figure 10 A wearable device 1000 including a computing environment 1002, a screen 1004, and a band 1006 is shown. The computing environment 1002 includes a first sensor 1008, a second sensor 1010, one or more processors 1012, and a memory 1014. It should be noted that Figure 10 The computing environment 1002 shown is intentionally simplified and, in some embodiments, can be closer to Figure 9 the computing system 900 shown. It should be noted that both the screen 1004 and the band 1006 are optional.

[0083] In one example, the first sensor 1008 includes a photoplethysmogram (PPG) sensor, and the second sensor 1010 includes an electrocardiogram (ECG) sensor. In an example embodiment, the computing environment 1002 includes only one of the first sensor 1008 or the second sensor 1010. In yet another example, the first sensor 1008 and / or the second sensor 1010 are separated from the computing environment 1002 and can be accessed through a separate interface or controller (such as Figure 9 the input controller 920 shown).

[0084] The first sensor 1008 and the second sensor 1010 are configured to obtain measurements or signals that can be received by one or more processors 1012, such as PPG signals and ECG sensors. One or more processors 1012 can be any general-purpose processor configured to execute a set of instructions. For example, one or more processors 1012 can be one or more general-purpose processors, one or more field-programmable gate arrays (FPGAs), and / or one or more application-specific integrated circuits (ASICs). In one example, one or more processors 1012 include one processor. Alternatively, one or more processors 1012 include multiple processors operatively connected. One or more processors 1012 are communicatively coupled to a memory 1014 via an address bus, a control bus, and a data bus (not shown).

[0085] The memory 1014 can be a random-access memory (RAM), a read-only memory (ROM), a persistent storage device such as a hard disk drive, an erasable programmable read-only memory (EPROM), etc. The memory 1014 can store information accessible by one or more processors 1012. For example, the memory 1014 (e.g., one or more non-transitory computer-readable storage media, memory devices) can include computer-readable instructions (not shown) executable by one or more processors 1012. The computer-readable instructions can be software written in any suitable programming language or can be implemented in hardware. Additionally, or optionally, the computer-readable instructions can be executed on one or more processors 1012 in logically and / or physically independent threads. For example, the memory 1014 can store instructions (not shown) that, when executed by one or more processors 1012, cause one or more processors 1012 to perform operations such as any operations and functions for which the wearable computing device 1000 is configured, as described herein. Further, or optionally, the memory 1014 can store data (not shown) that can be obtained, received, accessed, written, manipulated, created, and / or stored. For example, the data can include data and / or information described herein with respect to Figures 1 to 8 the description. In some implementations, the wearable computing device 1000 can obtain and / or store data from one or more memory devices remote from the wearable computing device 1000.

[0086] Figure 9 and Figure 10Example systems 900, 1000 are shown that can be used to implement aspects of the present invention. These computing systems are not intended to limit other computing systems, and various modifications of example systems 900, 1000 can also be used. The computing tasks discussed herein can be performed remotely from their respective systems and vice versa. Such configurations can be implemented without departing from the scope of the present invention. Various possible configurations, combinations, task divisions, and functional divisions can be made among components using a computer-based system. Computer-implemented operations can be performed on a single component or across multiple components. The tasks and / or operations performed by a computer can be executed sequentially or in parallel. Data and instructions can be stored in a single storage device or across multiple storage devices.

[0087] It should be noted that the foregoing description is only for enabling those skilled in the art to understand the present invention, and is not intended to limit the applicability of the present invention to other embodiments that can be easily understood and / or envisioned by the reader. In particular, although the present invention is mainly directed to the conversion of source (random) biological signals to target (random) biological signals, those skilled in the art will understand that the present invention is applicable to other technical fields involving the conversion of random signals having a common or related origin.

[0088] In the present invention, unless otherwise clearly stated or the context clearly indicates, items in the singular form should be understood to include items in the plural form and vice versa. Grammatical conjunctions are intended to express any and all disjunctive and conjunctive combinations of conjunctions, sentences, words, etc., unless the context otherwise indicates or clearly states. Thus, the term "or" should generally be understood to mean "and / or" etc. The use of any and all examples or exemplary language ("for example", "such as", "including", etc.) provided herein is only for better illustrating the embodiments and does not constitute a limitation on the scope of the embodiments or the claims.

[0089] The systems and methods of the present invention are applicable to any suitable programming language or environment, including but not limited to Java, C, C++, any suitable assembly language, Python, C#, JavaScript, Ruby, PHP, etc.

[0090] Some aspects described herein may relate to computer storage products having a non - transitory computer - readable medium (also referred to as a non - transitory processor - readable medium) having instructions or computer code thereon for performing various computer - implemented operations. The computer - readable medium (or processor - readable medium) is non - transitory because it does not itself include transitory propagated signals (e.g., propagated electromagnetic waves carrying information on a transmission medium such as space or a cable). The media and computer code (which may also be referred to as code) can be code designed and constructed for a particular purpose. Examples of non - transitory computer - readable media include, but are not limited to, magnetic storage media such as hard disks, floppy disks, and magnetic tapes; optical storage media such as CD / DVDs, CD - ROMs, and holographic devices; magneto - optical storage media such as optical disks; carrier - signal processing modules; and hardware devices specially configured to store and execute program code such as application - specific integrated circuits (ASICs), programmable logic devices (PLDs), read - only memories (ROMs), and random - access memory (RAM) devices. Other embodiments described herein relate to transient computer program products, which may include, for example, the instructions and / or computer code discussed herein.

[0091] Some of the processes / methods described herein may be executed by software (executed on hardware), hardware, or a combination thereof. Hardware modules include, for example, general - purpose processors, field - programmable gate arrays (FPGAs), and / or application - specific integrated circuits (ASICs). Software modules (executed on hardware) may be represented in various software languages (e.g., computer code), including C, C++, Java, Ruby, Visual Basic, Python, and / or other object - oriented, procedural, or other programming languages and development tools. Examples of computer code include, but are not limited to, microcode or microinstructions, machine instructions (such as those generated by a compiler), code for generating web services, and files containing higher - level instructions executed by a computer using an interpreter. For example, these processes / methods may be implemented using imperative programming languages (e.g., C, Fortran, etc.), functional programming languages (Haskell, Erlang, etc.), logic programming languages (e.g., Prolog), object - oriented programming languages (e.g., Java, C++, etc.), or other suitable programming languages and / or development tools. Other examples of computer code include, but are not limited to, control signals, encryption code, and compression code.

Claims

1. A method for converting a source signal into a target signal, the method comprising: deconstructing the source signal into a plurality of deconstructed source components; performing a transform on each of the plurality of deconstructed source components to generate a corresponding plurality of transformed source components; mapping a plurality of transformed source components to a plurality of estimated target components using a plurality of machine learning models, wherein each of the plurality of machine learning models is trained to map a corresponding one of the plurality of transformed source components to a corresponding one of the plurality of estimated target components; performing an inverse transform on each of the plurality of estimated target components to generate a corresponding plurality of transformed target components; and The target signal is generated by combining the plurality of transformed target components. The method of claim 1 , wherein the source signal and the target signal are biological signals of related origin. 3 . The method according to claim 1 , wherein the source signal is a photoplethysmogram (PPG) signal or an electrocardiogram (ECG) signal, and the target signal is an arterial blood pressure (ABP) signal.

4. The method according to claim 1 or 2, wherein the source signal comprises intervals of a predefined length obtained from a denoised input signal.

5. The method according to claim 1 or 2, wherein the source signal is deconstructed into the plurality of deconstructed source components using empirical mode decomposition.

6. The method of claim 1 or 2, wherein the transform comprises a Hilbert transform and the inverse transform comprises an inverse Hilbert transform.

7. An apparatus for cuffless blood pressure estimation, the apparatus comprising one or more processors and one or more non-transitory computer-readable media that collectively store instructions that, when executed by the one or more processors, cause the apparatus to perform operations comprising: Obtaining a source photoplethysmogram (PPG) signal; transforming a source PPG signal into a plurality of PPG components using a first transformation process; predicting a plurality of predicted arterial blood pressure (ABP) components from the plurality of PPG components using a plurality of prediction models, wherein each prediction model of the plurality of prediction models is operable to map from one or more of the plurality of PPG components to a predicted ABP component of the plurality of predicted ABP components; and The plurality of predicted ABP components are transformed into a target ABP signal using a second transformation process.

8. The apparatus of claim 7, wherein the operations further comprise: The target ABP signal is output.

9. The apparatus according to claim 7 or 8, wherein the second transform process is an inverse process of the first transform process.

10. The apparatus of claim 9, wherein the first transform process comprises a Hilbert-Huang transform and the second transform process comprises an inverse Hilbert-Huang transform.

11. The apparatus according to claim 7 or 8, wherein the source PPG signal comprises intervals of a predefined length obtained from an input PPG signal.

12. The apparatus of claim 11, wherein the operations further comprise: The input PPG signal is denoised before obtaining the source PPG signal.

13. The apparatus of claim 7 or 8, wherein the plurality of prediction models comprises a plurality of trained deep learning models.

14. The apparatus of claim 13, wherein the plurality of trained deep learning models comprises a plurality of trained U-net models.

15. The device according to claim 7 or 8, wherein the device is a wearable device.

16. A non-transitory machine-readable medium storing executable instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising: Decomposing a source random signal into a first decomposed component and a second decomposed component; generating a first component from the first decomposed components using a first transform function; generating a second component from the second decomposed components using a first transform function; mapping the first component to a first estimated component using a first trained machine learning model; mapping the second component to a second estimated component using a second trained machine learning model; generating a first target component from the first estimated component using a second transformation function, wherein the second transformation function is an inverse of the first transformation function; generating a second target component from the second estimated component using the second transformation function; and The first target component and the second target component are combined to reconstruct a target random signal.

17. The non-transitory machine-readable medium of claim 16, wherein mapping the first component to the first estimated component comprises providing both the first component and the second component as input to a first trained machine learning model.

18. The non-transitory machine-readable medium of any one of claims 16 to 17, wherein the operations further comprise: generating a third component from the source random signal; and mapping the third component to a third estimated component using a third trained machine learning model; The target random signal is generated by the first estimated component, the second estimated component, and the third estimated component.

Images