Physiological characteristic waveform classification with efficient deep network search

Abstract

A method for designing an artificial neural network includes performing a two-step neural architecture search, and selecting, for each respective layer, an operator that produces a maximum weighted output for each respective layer in the two-step neural architecture search. The two-step neural architecture search may include training a plurality of operators within each layer; and training a plurality of weights, each weight being applied to a respective operator combination within a respective layer, the method may be performed by executing instructions encoded on a non-transitory computer-readable memory with or without manual intervention based on resources of the processor.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] Priority and an earlier effective filing date are hereby claimed for all purposes, including priority rights, to U.S. Application Serial No. 63 / 354,502, filed on June 22, 2022. This related application is also hereby incorporated by reference herein for all purposes as if expressly set forth verbatim herein. Technical Field

[0003] The present disclosure relates to artificial neural networks, and more particularly, to neural architecture search that can be used to design artificial neural networks. Background Art

[0004] This section of this document presents information about and / or from the art that may provide context for or be relevant to the subject matter described herein and / or claimed below. It provides background information to facilitate a better understanding of various aspects of what is claimed below. This is a discussion of "related" art. Just because such art is related does not in any way mean that it is also "prior" art. Related art may or may not be prior art. The discussion in this section of this document should be read in this light, and not as an admission that it is prior art.

[0005] The physiological state of a patient in a clinical setting is often represented by monitoring one or more physiological characteristics of the patient over time. Monitoring often includes capturing data using sensors, the captured data representing each physiological characteristic. Because the data represents the physiological characteristics, the data may be referred to as "physiological data."

[0006] In a common example, called electrocardiography, multiple electrodes are placed at predetermined points on a patient's body. Each sensor generates a voltage signal that is captured and graphed. The graph, or electrocardiogram ("ECG" or "EKG"), graphically illustrates the magnitude of the voltage signal over time. A trained clinician or technician can then use the graphical representation to determine whether the patient's heart beat is "normal" or "abnormal." Thus, in this context, the captured physiological data represents the patient's heart beat.

[0007] The data captured by such monitoring is usually sorted by at least the time of acquisition and the attributes of the physiological characteristics represented in the data. In the ECG example above, the data is sorted not only by the time of acquisition and the voltage magnitude, but also by the sensor that acquired the data. The voltage acquired by the individual sensors can be presented to human perception as a graph or plot of the voltage magnitude over time, called a "waveform." Physiological data is also commonly referred to collectively as a "waveform." The result of the ECG, then, is a set of waveforms that reflect the patient's heart beats.

[0008] As mentioned above, with respect to the electrocardiogram, a trained clinician or technician can analyze the "waveform" to determine certain aspects of the patient's condition captured in the physiological data. For a variety of reasons, technology has turned to "smart machines," or computing machines that employ various forms of artificial intelligence, to perform this analysis or assessment. However, like a clinician or technician, a smart machine must be trained to perform this analysis.

[0009] One type of artificial intelligence is a "neural network", sometimes also called an "artificial neural network" or "ANN". A technique sometimes used to train intelligent machines (and more specifically artificial neural networks) is "neural architecture search" or "NAS". NAS is a technique for automating the design of artificial neural networks. NAS has been used to design ANNs that are comparable to or better than hand-designed ANNs.

[0010] Artificial neural networks can be large and complex, so designing them can consume a lot of time and effort. Even automated NAS can take a lot of time and computing resources. For example, NAS can consume anywhere between 100 and several thousand processor hours. Result accuracy is also important, and all else being equal, higher result accuracy is generally better. But be aware that the design process can involve trade-offs. Therefore, the technology is open to new NAS methods that improve result accuracy, computing resource consumption, and design time. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Illustrative embodiments of the subject matter claimed below will now be disclosed. For the sake of clarity, not all features of actual implementations are described in this specification. It should be understood that in the development of any such actual implementation, many implementation-specific decisions must be made to achieve the developer's specific goals, such as complying with system-related and business-related constraints, which will vary from implementation to implementation. In addition, it should be understood that such development work, even if complex and time-consuming, will be routine work for those of ordinary skill in the art who benefit from this disclosure.

[0012] Figure 1 An artificial neural network designed according to an embodiment of the present disclosure is conceptually depicted.

[0013] Figure 2A-2B Illustration for design Figure 1 A specific embodiment of the artificial neural network method.

[0014] Figure 3 A specific embodiment of a computing device is schematically depicted, by which the computing device can be used in Figure 1 The design of artificial neural networks is implemented in Figure 2A-2B method.

[0015] Figure 4A-4B A specific embodiment of a method for designing an artificial neural network according to the present disclosure is illustrated.

[0016] Figure 5 Conceptual depiction of Figure 4A-4B Artificial neural network designed by this method.

[0017] Figure 6 , Figure 7A-7E and Figure 8A-B The diagram shows the Figure 1 Artificial neural network or Figure 5 Various operators of artificial neural networks.

[0018] Fig. 9 A specific embodiment of a computing device is schematically depicted by which a computing device can be used in a design Figure 5 The artificial neural network is executed Figure 4A-4B method.

[0019] Fig.10 Presented by Fig. 9 The computing device executes the workflow to achieve Figure 4A-4B Methods to design Figure 5 of artificial neural networks.

[0020] Fig.11 An ECG acquisition of an input sample beat is illustrated.

[0021] Fig.12 Represents the acquired ECG waveform.

[0022] Figure 13A-13B Example N beats are presented.

[0023] Figure 14A-14B Example S beats are presented.

[0024] Figure 15A-15B An example B beat is presented.

[0025] Figure 16A-16B An example V beat is presented.

[0026] Figure 17A-17B An example F beat is presented.

[0027] Figure 18A-18B Example Q beats are presented.

[0028] Although the implementation of the present disclosure is suitable for various modifications and alternative forms, the accompanying drawings illustrate the specific examples described in detail herein by way of example. However, it should be understood that the description of the specific examples herein is not intended to limit the present disclosure to the specific forms disclosed, but rather, the inventive concepts herein are intended to cover all modifications, equivalents and alternatives falling within the spirit and scope of the present disclosure defined by the appended claims. DETAILED DESCRIPTION

[0029] The present disclosure provides a NAS technique that produces excellent result accuracy and can also reduce computing resource consumption and design time relative to many conventional techniques. Therefore, the disclosed NAS technique improves the operational performance of computing resources used to design and operate the underlying ANN. In addition, the disclosed NAS technique is embedded in practical applications - namely the design of ANNs. In addition, NAS can be used to design ANNs that are optimized to analyze certain outputs. One such embodiment is discussed further below.

[0030] In one embodiment, a method for designing an artificial neural network includes: performing a two-step neural architecture search, and selecting an operator for each respective layer that produces a maximum weighted output for each respective layer in the two-step neural architecture search. The two-step neural architecture search may include: training a plurality of operators within each layer; and training a plurality of weights, each weight being applied to a respective combination of operators within a respective layer.

[0031] In another embodiment, a method for designing an artificial neural network includes: creating a deep neural network including multiple layers, each layer including multiple operators; accessing multiple sample inputs; performing a two-step neural architecture search using the accessed sample inputs; and, selecting an operator for each corresponding layer that produces a maximum weighted output for each corresponding layer in the two-step neural architecture search. The two-step neural architecture search may include: training multiple operators within each layer; and training multiple weights, each weight being applied to a corresponding combination of operators within a corresponding layer.

[0032] In yet another embodiment, a method for creating a deep neural network includes: defining a deep neural network; obtaining a plurality of sample inputs; and training the deep neural network on the sample inputs. Defining the deep neural network may include: defining a plurality of layers, each layer receiving a layer input and generating a layer output, each layer including a plurality of operators, each operator including a plurality of operator parameters; and for each of the plurality of operators within each layer, pairing each operator with an instance of itself and an instance of each other operator within the layer, each instance of each operator receiving a layer input from a previous layer and generating an operator output; summing the operator outputs of each pair of operators; and weighting the summed operator outputs, the sum of the weights being equal to 1, to generate a plurality of layer outputs. Training the deep neural network may include: a first round in which each weight is fixed to train the operator parameters; a second round in which the operator parameters are fixed to train the weights; selecting for each layer a pairing operator that produces the maximum weighted output to provide the layer output; and deselecting for each layer each pairing operator that does not produce the maximum weighted output.

[0033] In yet other embodiments, a computing device includes: a processor-based resource; and a memory in electronic communication with the processor-based resource. The memory may be encoded with instructions that, when executed by the processor-based resource, perform the methods described herein.

[0034] Still other embodiments include a non-transitory computer-readable medium encoded with instructions that, when executed by a processor-based resource, perform the methods described herein.

[0035] One or more specific embodiments of the subject matter currently claimed will be described below in conjunction with the drawings herein. The subject matter currently claimed is not limited to the embodiments and illustrations contained herein, but includes modifications of these embodiments, including combinations of parts of the embodiments and elements of different embodiments, which modifications fall within the scope of the appended claims. In the development of any such actual implementation, as in any engineering or design project, many implementation-specific decisions must be made to achieve the developer's specific goals, such as complying with system-related and business-related constraints, which may vary from implementation to implementation. In addition, such development work may be complex and time-consuming, but for ordinary technicians who benefit from this disclosure, it will still be a routine task of design, manufacturing, and processing.

[0036] Turning now to the accompanying drawings, Figure 1 An artificial neural network 100 designed according to an embodiment of the present disclosure is conceptually depicted, and Figure 2A-2B Illustration for design Figure 1This particular embodiment assumes that the artificial neural network 100 has been designed in its broad form, i.e., Figure 1 Thus, the number of layers, the number of operators per layer, the identity of the layers, etc. have been determined.

[0037] More specifically, the artificial neural network 100 includes Z layers, designated as layer 1-layer Z. Each layer includes n operators, designated as O1-On. Within each layer, each operator is paired with itself and each other operator. Thus, operator O1 is paired with itself and each other operator O2-On, resulting in m pairs of operators (note that m=n!+n). Each of the m pairs in each layer is individually weighted. In Figure 1 In the nomenclature, the weight w 11 is the first weight in the first layer, weight w 12 is the second weight in the first layer, and so on. In this particular embodiment, the sum of the m weights in each layer is equal to 1.

[0038] The information flow in an artificial neural network 100 is Figure 1 100 from left to right, as indicated by the arrows therein. Each of the Z layers has an input and an output. The input to each layer, except for layer 1, is the output of the immediately preceding layer. The input to layer 1 includes a plurality of samples, the nature and identity of which will depend on the ultimate use of the artificial neural network 100. The output of layer Z is the classification of any given input sample that has just been processed by the artificial neural network 100.

[0039] The neural architecture search disclosed herein determines which of the m pairs of operators will be selected in each of the Z layers for the design of the artificial neural network 100. One embodiment of this technique is Figure 2A-2B The method 200 is used to design an artificial neural network, such as Figure 1 The method 200 begins (at 205) by performing a two-step neural architecture search, such as Figure 2B The two-step neural architecture search (at 205) may include: Figure 2B As shown in: training (at 207) a plurality of operators within each layer and training (at 208) a plurality of weights, each weight being applied to a corresponding combination of operators within a corresponding layer. Return to Figure 2A , method 200 continues (at 210) by selecting an operator for each respective layer that produces a maximum weighted output for each respective layer in a two-step neural architecture search.

[0040] As mentioned above, neural architecture search is an automated process. Figure 3 A specific embodiment of a computing device 300 is schematically depicted, by which computing device 300 can be used in Figure 1 The design of artificial neural networks is implemented in Figure 2A-2B The computing device 300 may be, for example, a workstation. In this particular embodiment, the computing device 300 generally includes a user interface 310, a processor-based resource 320, and a memory 330. The user interface 320 includes a display 311, one or more peripheral devices (such as a keyboard 312 and / or a mouse 313), and a software component ("UI") 314 resident on the memory 330. The UI software component 314 is executed by the processor-based resource 320 to provide a presentation (not separately shown) on the display 311, and the user 340 can interact with the presentation using (one or more) peripheral components.

[0041] Processor-based resources 320 execute machine-executable instructions 324 residing in memory 330 to perform the functions of the techniques described herein. Instructions 324 may be embedded in memory 330 as firmware, or encoded as routines, subroutines, applications, etc. Memory 330 is a non-transitory computer-readable medium that is readable by a computer and may be local or remote.

[0042] Thus, in the illustrated embodiment, computing device 300 performs software-implemented functions of the presently disclosed technology. More specifically, processor-based resources 320 execute instructions 324 (all in Figure 3 310 ) to perform the programming functions of the disclosed neural architecture search technology disclosed herein. As part of the artificial neural network 100 design process, the user 340 can call the neural architecture search technology through the user interface 310.

[0043] The neural architecture search described above can be set in the broader context of designing entire artificial neural networks. To this end, Figure 4A-4B A method 400 for designing an artificial neural network is illustrated. For illustrative purposes, Figure 5 Method 400 is discussed in the context of an artificial neural network 500 shown in . Artificial neural network 500 is of a type known as a deep neural network, and will be referred to as such throughout the rest of this article.

[0044] Method 400 begins by creating (at 405) a deep neural network, such as deep neural network 500. Deep neural network 500 includes a plurality of layers, each layer including a plurality of operators. More specifically, deep neural network 500 includes three layers, each of layers 1 to 3 also including three operators O1-O3. Note that in each layer, each operator is paired with itself and with each of the other two operators. Thus, each of layers 1 to 3 includes 12 operator pairs. Each pair is summed, weighted, and transmitted as part of the output of the corresponding layer.

[0045] More technically, each layer is a two-branch structure that maps from an input tensor (e.g., the output of the previous layer) to an output tensor. This can be described as input, O1, O2, add. Add is the operator that combines the individual outputs to form the output tensor of the block. Figure 5 The operator pairs shown in are all candidates to become blocks that become layers in the final design of the artificial neural network. The two-step neural architecture search disclosed herein determines which candidate is selected for the final design.

[0046] A skilled person having the benefit of this disclosure will appreciate that the design of the artificial neural network in any given embodiment may be implementation specific. Implementations of artificial neural networks may differ in both the number of layers and the number of operators. In one embodiment, for example, the artificial neural network includes eight layers, each layer including 13 operators, resulting in 13! +13 pairs. Using the nomenclature established above, in this particular implementation, Z = 8, n = 13, and m = 13! +13.

[0047] Operators O1-ON may be selected from any of a variety of operators. Figure 6 , Figure 7A-7E and Figure 8A-B The diagram shows the Figure 1 and / or Figure 5 The various operators of artificial neural networks. The parameters of a filter are the weights, weight layers, or coefficients of the filter. Figure 6 Illustration of the application of a rectified linear unit ("ReLU"). Figure 7A-7E Operations including one or more rectified linear units (“ReLU”) and batch normalization (“BN”) in different combinations and orders are illustrated. Figure 8A-B Two implementations of a variant of the operator called ResNet, which is referred to in the art as a residual network ("ResNet"). ResNet is a type of convolutional neural network, which is a type of deep neural network. Figure 8A-B The implementation shown in can use different parameters (such as the number of filters). For example, the number of filters can be 32, 64 or 128 instead of 256. Other operators known in the art (such as "Maxpooling") can also be used.

[0048] A skilled person having the benefit of this disclosure will appreciate that the choice of operators, their order, etc. will depend on the use of the artificial neural network. In some embodiments, the artificial neural network 100, 500 is used in a clinical setting to classify waveforms representing physiological characteristics of a patient. This use will be discussed further below.

[0049] Back to Figure 4A, after creating (at 405) the deep neural network, method 400 then proceeds to access (at 410) a plurality of sample inputs. In the illustrated embodiment, the sample inputs are "labeled inputs," and thus, the training discussed below is a form of what is known as "supervised learning." For example, as will be discussed further below, the sample inputs are waveforms representing individual heart beats of a patient. These are sometimes referred to as "beats" in the corresponding art. These beats are classified as set forth in Table 1 below, and the classification is associated with the sample inputs to "label" the sample inputs.

[0050] Table 1: Pulse classification

[0051]

[0052] As described above, in a specific embodiment not shown, the deep neural network includes eight layers, each layer includes 13 operators, which become 13! + 13 sum pairings. In this specific embodiment, the number of labeled samples is as described in Table 2 below. In this specific embodiment, the length of each labeled sample is 133 sample points. Of these labeled samples, 75% are randomly selected and used for training, while the remaining 25% are used for validation.

[0053] Table 2. Number of samples by category

[0054] Classification quantity N 66,666 V 61,922 F 3,970 G 56,638 Q 27,582

[0055] Now unified reference Figure 4A-4B , and then using the sample input accessed (at 410) to execute (at 415) Figure 4B The two-step neural architecture search shown in . However, as a preliminary matter, this particular embodiment first reuses the continuous relaxation function represented by equations (1)-(2). In equations (1)-(2), k≡ the number of weights or operators, and m≡ the total number of pairs of operators in a given layer.

[0056]

[0057] Performing (at 415) a two-step neural architecture search using the accessed (at 410) sample inputs includes training (at 417) a plurality of operators within each layer and training (at 418) a plurality of weights. Each weight is applied to a corresponding combination of operators within a corresponding layer. More specifically, the weights are fixed while the operators are trained (at 417) using 75% of the sample inputs selected at random. Then, the weights are trained (at 418) using the same 75% of the sample inputs selected at random while the operators are fixed. Note that the order of training (at 417) the operators and training (at 418) the weights is not important with respect to each other.

[0058] Reference again Figure 4A , method 400 then selects (at 420) for each respective layer the operator that produced the maximum weighted output for each respective layer in the two-step neural architecture search. This will typically include deselecting other paired operators that do not produce the maximum weighted output. More technically, once training (at 415) is complete, the network architecture can be finalized by retaining the strongest predecessors (with strengths according to weights) for each block, and then selecting the most likely operator by taking "argmax" - i.e., determining the block that gives the maximum value for the respective layer. Once the network architecture is finalized, the artificial neural network can be validated using 25% of the input samples that were not used in training.

[0059] As mentioned above, neural architecture search is an automated process. Fig. 9 A specific embodiment of a computing device 900 is schematically depicted, by which computing device 900 can be used to Figure 5 The design of artificial neural networks is implemented in Figure 4A-4B The computing device 900 may be, for example, a workstation. In this particular embodiment, the computing device 900 generally includes a user interface 910, a processor-based resource 920, and a memory 930. The user interface 920 includes a display 911, one or more peripheral devices (such as a keyboard 912 and / or a mouse 913), and a software component ("UI") 914 resident on the memory 930. The UI software component 914 is executed by the processor-based resource 920 to provide a presentation (not shown separately) on the display 911, and the user 940 can interact with the presentation using (one or more) peripheral components.

[0060] The processor-based resource 920 executes machine executable instructions 924 residing in the memory 930 to perform the functions of the technology described herein. The instructions 924 can be embedded in the memory 930 as firmware, or encoded as routines, subroutines, applications, etc. The memory 930 is a computer-readable non-transitory computer-readable medium and can be local or remote. In addition, a labeled input sample library 931, an operator library 932, and a deep neural network 933 reside in the memory 930.

[0061] Thus, in the illustrated embodiment, the computing device 900 performs the software-implemented functionality of the presently disclosed technology. More specifically, processor-based resources 920 execute instructions 924 (each such as Fig. 9) to perform the programming functions of the disclosed neural architecture search techniques disclosed herein and the design of the deep neural network 933. The design of the deep neural network 933 can be performed using input from a user 940 through the user interface 910. As part of the design process of the artificial neural network 100, the neural architecture search technique can be called by the user 940 through the user interface 910.

[0062] Now unified reference Figure 5 and Fig. 9 In one particular embodiment, processor-based resources 920 execute instructions 924 to perform Fig.10 The workflow 1000 shown in FIG. The workflow 1000 starts from Figure 5 The initial architecture in begins with the initial design (at 1003) of an artificial neural network ("ANN") 500. As described above, the artificial neural network 500 is a deep neural network, and will be referred to as such throughout the rest of this article.

[0063] The initial design begins (at 1003) by receiving (at 1006) the number of layers and receiving (at 1006) the identities of operators O1-O3 (at 1009). This can be done as follows: Fig. 9 The user interface 910 in receives input from the user 940. The operators O1-O3 may be identified in some way based on the contents of the operator library 932 presented to the user 940 through the user interface 910. Alternatively, they may be identified by manual input or definition by the user 940. The processor-based resource 920 executing the instructions 924 then fleshes out (at 1012) the initial architecture of the deep neural network 500, such as Figure 5 as shown in .

[0064] Note that in some embodiments, the initial design of the deep neural network 500 (at 1003) can be performed in a more automated manner. For example, the definition of the initial architecture can be stored in an electronic file (not shown) residing in the memory 930. The processor-based resource 920 can access the electronic file to receive the initial definition of the architecture when executing the instructions 924. The processor-based resource 920 can then access the identified operators from the operator library 932 and enrich the initial architecture.

[0065] The processor-based resource 920 then begins (at 1015) an artificial neural network design. This can occur automatically without human intervention after the initial design is completed (at 1003), or be invoked by a user 940 through the user interface 910. The artificial neural network design (at 1015) begins by accessing (at 1018) the labeled samples 931. The accessed (at 1018) samples are then used to perform (1021) a two-step neural architecture search as described above. This typically involves first training multiple operators within each layer with weights fixed, and then training the weights.

[0066] Then, the processor-based resources 920 finalize (at 1024) the architecture. The finalization (at 1024) generally includes: selecting for each corresponding layer the operator that produces the maximum weighted output for each corresponding layer in the two-step neural architecture search, while deselecting other paired operators that do not produce the maximum weighted output. Once the network architecture is finalized (at 1024), the artificial neural network is verified (at 1027), both of which are also described above.

[0067] The artificial neural network designed using the two-step neural network shows improved performance in terms of result accuracy relative to conventional methods. A specific embodiment is mentioned above, in which the artificial neural network computing resource consumption and design time include eight layers, each layer includes 13 operators, which become 13! + 13 summations and weighted pairings. In this specific embodiment, the input samples are labeled with the classifications listed in Table 1, and the number of labeled samples in each classification is listed in Table 2. The length of each labeled sample in this specific embodiment is 133 sample points. In this embodiment, after 100 epochs, the training accuracy is approximately 97.7%, and the verification accuracy is approximately 96.8%.

[0068] The technology disclosed herein assumes that a set of input samples have been previously collected. The input samples in the disclosed embodiments are ECG waveforms. However, it should be understood that this is for illustration only, and the embodiments can also operate on input samples of waveforms collected using other processes. For example, in patient care, it is common to monitor many physiological characteristics of patients, which can be collected and represented as waveforms using other programs. The technology described herein can be used in combination with other waveforms collected by other programs and representing other physiological characteristics. Examples of such other physiological characteristics include, but are not limited to, invasive pressure (e.g., invasive blood pressure), gas output (e.g., oxygen, carbon dioxide) from breathing, blood oxygen, internal pressure (e.g., intracranial pressure), anesthetic (e.g., nitrous oxide) concentration measurements, etc.

[0069] Furthermore, the input samples disclosed herein are "organic" in the sense that they are collected by actually performing an ECG procedure on (one or more) patients. Input samples in other embodiments may be synthetic in the sense that they are collected by artificially generating them. Still other embodiments may use a combination of organic and synthetic waveforms.

[0070] ECG waveforms are collected over time and represent the number of heart beats that occur within a predetermined time window. These waveforms can be sampled as several parts of a larger ECG waveform, each part representing a single heart beat. Therefore, the ECG waveforms in the input samples of the illustrated embodiment can be referred to as "beats" because they are several parts of a larger waveform representing a single heart beat. Therefore, the input samples of the disclosed technology can be obtained by sampling several parts of a larger waveform.

[0071] As described above, the input ECG samples are beats labeled with the classifications listed in Table 1. One criterion for selecting the input sample set is that it contains a statistically significant number of each potential outcome. Thus, in the illustrated embodiment, there should be a statistically significant number of N, B, S, V, F, and Q beats. As described above, a beat is a waveform that is actually an ordered data set.

[0072] However, please note that other embodiments may be used to analyze and / or classify input samples representing some other physiological characteristics. Examples of such other physiological characteristics include, but are not limited to, invasive pressures (e.g., invasive blood pressure), gas outputs from breathing (e.g., oxygen, carbon dioxide), blood oxygen, internal pressures (e.g., intracranial pressure), anesthetic (e.g., nitrous oxide) concentration measurements, and the like. Therefore, in some embodiments, the classification of input samples may vary to accommodate various aspects of the physiological characteristics of interest to clinicians. The number of classifications may also vary, from as few as two to more than the six seen herein. Still other physiological characteristics and classifications may become apparent to those skilled in the art having the benefit of this disclosure.

[0073] Although the disclosed techniques assume that input samples have been previously acquired, for the sake of completeness, acquiring input samples in an ECG context will now be discussed. Fig.11 An ECG procedure 1100 according to the present disclosure is illustrated. Fig.11 In FIG. 1 , a patient 1103 is undergoing an ECG procedure 1100 performed using an ECG system 1106. The ECG system 1106 includes an ECG monitor 1109, a plurality of electrical leads 1112, and a plurality of ECG electrodes 1115 (only one is shown). The ECG electrodes 1115 and the electrical leads 1112 are not necessarily connected as shown in FIG. Fig.11A common configuration and one that can be used to practice the presently disclosed technology is a 10 electrode, 12 lead configuration for measuring 10 voltages across the body.

[0074] The ECG system 1106 acquires a plurality of ECG waveforms 1120, such as Figure 1 120. The ECG waveform 1120 is then processed to generate individual input samples or beats 1122. The processing may be performed by the ECG monitor 109. However, more likely, the ECG waveform 1120 will be exported to another computing device or computing system that performs the processing. The processing may be performed manually, for example, on a workstation. In other embodiments, the processing may be performed automatically by a neural network (not shown) trained for beat detection in ECG waveforms. Likewise, Fig.11 11. ECG waveform 1120 and input samples 1122 are shown, presented for human perception.

[0075] Once collected, the sample pulse is labeled. The practice of the technology disclosed herein is independent of the manner in which the input sample is labeled. In some embodiments, the input sample pulse can be used by, for example, user 940 Fig. 9 The user 940 may be manually labeled by the computing device 900 in FIG. Depending on the task being performed and the skills involved, the user 940 may be a person who is trained and skilled in classifying input samples. For example, in embodiments where the input sample is a beat (such as in this disclosure herein), the user 940 may be a clinician or technician who is capable of classifying and labeling the input sample as described above.

[0076] The N, B, S, V, F, and Q beat categories are defined in Table 1. Each beat example is Figures 13A-18B More specifically, Figure 13A-13B An example N beat is presented; Figure 14A-14B An example B beat is presented; Figure 15A-15B An example S beat is presented; Figure 16A-16B An example V beat is presented; Figure 17A-17B An example F beat is presented; and Figure 18A-18B Example Q beats are presented.

[0077] Once an artificial neural network is designed using the two-step neural architecture search disclosed herein, it can be deployed to classify inputs that are similar in nature to the inputs it was trained on. For example, the disclosed embodiments were trained on ECG samples classified as N, B, S, V, F, and Q beats, and they can be used to classify additional beats as N, B, S, V, F, and Q beats in a clinical setting. The improved output accuracy of the disclosed embodiments will directly translate to the clinical setting.

[0078] Therefore, the neural architecture search disclosed herein improves the prior art of artificial neural network design and implementation. This improvement translates into the prior art regardless of the purpose for which the trained artificial neural network is used. Furthermore, these are practical applications of the two-step neural architecture search that is at the heart of the disclosed embodiments.

[0079] The above discussion of computing technologies related to the disclosed technology includes repeated references to one or more "processor-based resources." Examples include Figure 3 The processor-based resources 320 and Fig. 9 The processor-based resources 920 in FIG. As will be understood by those skilled in the art who have the benefit of this disclosure, the term "processor" is understood in the art to have a clear structural connotation. A processor may be hardware, software, or some combination of the two.

[0080] exist Figure 3 and Fig. 9 In the illustrated embodiment, the processor-based resources 320, 920 are programmed hardware processors, such as controllers, microcontrollers, or central processing units ("CPUs"). However, in alternative embodiments, the processor-based resources 320, 920 may be digital signal processors ("DSPs"), graphics processors, processor chipsets, application-specific integrated circuits ("ASICs"), appropriately programmed electrically programmable read-only memories ("EPROMs"), appropriately programmed electrically erasable programmable read-only memories ("EEPROMs"), or any other equivalent integrated or discrete logic circuits, as well as any combination of these components and one or more machine learning algorithms. Other implementations of the processor-based resource(s) of the computing techniques disclosed herein may be appreciated by those skilled in the art.

[0081] Similarly, the above discussion of computer technology includes references to "non-transitory computer-readable media," more colloquially referred to in some places and by those skilled in the art as "memory." Examples include Figure 3 The memory 330 and Fig. 9 930 in. The memory 330, 930 may be distributed, for example, in a computing cloud. The memory 330, 930 may include read-only memory ("ROM"), random access memory ("RAM"), or a combination of the two. The memory 330, 930 will typically be installed memory, but may be removable. The memory 330, 930 may be a primary storage device, a secondary storage device, a tertiary storage device, or some combination thereof implemented using electromagnetic, optical, or solid-state technology. Thus, in various embodiments, the memory 330, 930 may be part of a mass storage device, a hard drive, a solid-state drive, an external drive (whether magnetic or solid-state), an optical disk, a magnetic disk, a portable external drive, a jump drive, and the like.

[0082] The computing techniques disclosed above are also all shown as a single device located in a single computing device. However, this is only for ease of illustration. A skilled person who benefits from this disclosure will recognize that various aspects of the computing techniques can be distributed. The memory 330, 530 and the processor-based resources 310, 910 can be distributed, for example, in a computing cloud. For another example, in some embodiments, Figure 3 The computing device 300 and / or Fig. 9 The computing device 900 in the embodiment may be a personal computing device other than a workstation. The computing device 300 and the computing device 900 may also access data through a network such as a local area network ("LAN").

[0083] In other embodiments, computing device 300 and / or computing device 900 may be a computing system. In such a computing system, processing and memory resources may be distributed in a cloud, for example, accessed from a workstation or other personal computing device via a public or private network such as the Internet. Still other variations of computing device 300 and / or computing device 900 may be implemented by those skilled in the art having the benefit of the present disclosure.

[0084] The features of several embodiments are summarized above so that those of ordinary skill in the art can better understand the various aspects of the present disclosure. Those of ordinary skill in the art should understand that they can easily use the present disclosure as a basis for designing or modifying other processes and structures to implement the same purposes and / or achieve the same advantages as the various embodiments introduced herein. Those of ordinary skill in the art should also recognize that such equivalent constructions do not depart from the spirit and scope of the present disclosure, and they can make various changes, substitutions and modifications to this article without departing from the spirit and scope of the present disclosure.

[0085] Although the subject matter has been described in language specific to structural features or methodological acts, it should be understood that the subject matter of the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing at least some of the claims.

[0086] Various operations of the embodiments are provided herein. The order of describing some or all operations should not be interpreted as implying that these operations must rely on the order. In the case of benefiting from this description, alternative orders will be understood. In addition, it should be understood that not all operations are necessarily present in each embodiment provided herein. In addition, it should be understood that not all operations are necessary in some embodiments.

[0087] It should be understood that the layers, features, elements, etc. depicted herein are illustrated with specific dimensions relative to each other (such as structural dimensions or orientations), for example, for simplicity and ease of understanding, and in some embodiments, their actual dimensions are very different from the dimensions shown herein. In addition, "exemplary" is used herein to indicate use as an example, instance, illustration, etc., without necessarily having an advantage. As used in this application, "or" is intended to mean an inclusive "or" rather than an exclusive "or". In addition, "one" and "a" used in this application and the appended claims should generally be interpreted as meaning "one or more", unless otherwise specified or it is clear from the context that it refers to the singular form.

[0088] In addition, at least one of A and B and / or similar words generally represent A or B or both A and B. In addition, to the extent that "includes," "has," "contains," "with," or variations thereof are used, these terms are intended to be inclusive in a manner similar to the term "comprising." In addition, unless otherwise specified, "first," "second," or similar words are not intended to imply temporal aspects, spatial aspects, order, etc. Instead, these terms are used only as identifiers, names, etc. of features, elements, items, etc. For example, a first element and a second element generally correspond to an element A and an element B or two different or two identical elements or identical elements.

[0089] Furthermore, the phrase "capable" as used herein is an acknowledgement of the fact that some of the functions described for various parts of the disclosed apparatus are only performed when the apparatus is powered on and / or operating. A skilled artisan having the benefit of this disclosure will appreciate that the embodiments described herein include many electronic or electromechanical parts that require power to operate. Even when power is provided, some of the functions described herein only occur when in operation. Thus, sometimes, some embodiments of the apparatus of the presently claimed subject matter are "capable" of performing the functions described, even when they are not actually performing those functions - i.e., when there is no power or when they are powered on but not operating.

[0090] This concludes the detailed description. The particular examples disclosed above are illustrative only, as the presently claimed subject matter may be modified and practiced in different but equivalent manners, as will be apparent to those skilled in the art having the benefit of the teachings herein. Furthermore, no limitation is intended to the details of construction or design shown herein, except as described in the claims below. It is evident, therefore, that the particular examples disclosed above may be altered or modified, and all such variations are considered to be within the scope and spirit of the presently claimed subject matter. The protection sought herein, therefore, is as described in the claims below.

Claims

1. A method for designing an artificial neural network, comprising: Perform a two-step neural architecture search consisting of: Training multiple operators within each layer; and training a plurality of weights, each weight being applied to a corresponding combination of operators within a corresponding layer; and The operator is selected for each corresponding layer, the operator producing a maximum weighted output for each corresponding layer in the two-step neural architecture search.

2. The method according to claim 1, further comprising: Create a deep neural network consisting of multiple layers, each layer consisting of multiple operators; accessing a plurality of sample inputs, performing the two-step neural architecture search using the sample inputs, The plurality of sample inputs are recorded electrocardiogram waveforms.

3. A computing device, comprising: Processor-based resources; as well as A memory in electronic communication with the processor-based resource and encoded with instructions that, when executed by the processor-based resource, perform the method of any one of claims 1 to 2.

4. A non-transitory computer readable medium encoded with instructions that, when executed by a processor-based resource, perform the method of any one of claims 1 to 2.

5. A method for designing an artificial neural network, comprising: Create a deep neural network consisting of multiple layers, each layer consisting of multiple operators; Access multiple sample inputs; Performing a two-step neural architecture search using the accessed sample input, the two-step neural architecture search comprising: Training multiple operators within each layer; and training a plurality of weights, each weight being applied to a corresponding combination of operators within a corresponding layer; and The operator is selected for each corresponding layer, the operator producing a maximum weighted output for each corresponding layer in the two-step neural architecture search. The method according to claim 5 , wherein the number of layers is three and the number of operators is three.

7. The method of claim 5, wherein the number of layers is eight and the number of operators is thirteen.

8. The method of claim 5, wherein creating a deep neural network comprises: receiving definitions of the plurality of layers, each layer receiving a layer input and generating a layer output, each layer comprising a plurality of operators, each operator comprising a plurality of operator parameters; For each of the multiple operators within each layer: pairing each operator with its own instance and an instance of each other operator within the layer, each instance of each operator receiving a layer input from a previous layer and generating an operator output; Sum the operator outputs for each pair of operators; as well as The summed operator outputs are weighted so that the sum of the weights equals 1 to generate multiple layer outputs; 9. The method of claim 5, wherein the samples are recorded waveforms representing physiological characteristics of a human body.

10. The method of claim 9, wherein the recorded waveform is an electrocardiogram beat.

11. The method of claim 5, wherein the first step of training the plurality of operators within each layer comprises fixing each weight to train the operator.

12. The method according to claim 11, wherein: Each operator includes a plurality of operator parameters; and The second step of training the weights includes fixing the operator parameters, including fixing the operator parameters to train the weights.

13. The method according to claim 5, wherein: Each operator includes a plurality of operator parameters; and The second step of training the weights includes fixing the operator parameters, including fixing the operator parameters to train the weights.

14. The method of claim 5, wherein training the deep neural network on the sample inputs comprises, for each layer, deselecting each set of pairwise operators that does not produce a maximum weighted output.

15. The method of claim 5, wherein each layer is a dual-branch structure mapping from an input tensor to an output tensor.

16. A computing device comprising: Processor-based resources; as well as A memory in electronic communication with the processor-based resource and encoded with instructions that, when executed by the processor-based resource, perform the method of any one of claims 5 to 15.

17. A non-transitory computer readable medium encoded with instructions that, when executed by a processor-based resource, perform the method of any one of claims 5 to 15.

18. A method for creating a deep neural network, the method comprising: Define a deep neural network, including: defining a plurality of layers, each layer receiving a layer input and generating a layer output, each layer including a plurality of operators, each operator including a plurality of operator parameters; For each of the multiple operators within each layer: pairing each operator with an instance of itself and an instance of each other operator within the layer, each instance of each operator receiving a layer input from a previous layer and generating an operator output; summing the operator outputs for each pair of operators; and The summed operator outputs are weighted so that the sum of the weights equals 1 to generate multiple layer outputs; obtaining a plurality of sample inputs; and Training the deep neural network on the sample input, the training comprising: The first round, where each weight is fixed to train the operator parameters; and the second round, where the operator parameters are fixed to train the weights; selecting for each layer the pairing operator that produces the maximum weighted output to provide the layer output; and For each layer, deselect each set of pairwise operators that does not produce the maximum weighted output. The method of claim 18 , wherein the number of layers is three and the number of operators is three.

20. The method of claim 18, wherein the number of layers is eight and the number of operators is 13.

21. The method of claim 18, wherein the samples are recorded waveforms representing physiological characteristics of a human body.

22. The method of claim 21, wherein the recorded waveform is an electrocardiogram beat.

23. A computing device comprising: Processor-based resources; and A memory in electronic communication with the processor-based resource and encoded with instructions that, when executed by the processor-based resource, perform the method of any one of claims 18 to 22.

24. A non-transitory computer readable medium encoded with instructions that, when executed by a processor-based resource, perform the method of any one of claims 18 to 22.

Images