Feature sorting neural network and method, method for generating simplified feature set model

Abstract

This approach involves feature ranking within deep neural networks in multi-task and / or multi-label settings. Methods for identifying task-specific features as well as features shared by multiple tasks are described. In addition to facilitating interpretability, the selected feature subsets can be used to craft efficient models that achieve better stability and regularization as well as reduced computational and memory usage.

Description

[0001] Cross-application of related applications

[0002] This application claims priority to U.S. Patent Application No. 15 / 799,698, filed on October 31, 2017. Background Art

[0003] The subject matter disclosed herein relates to feature ranking using neural networks, and more particularly to the process of feature ranking using machine learning techniques to facilitate the selection of features of interest.

[0004] Deep learning algorithms, such as artificial neural networks, are learning tools for various data analysis tasks such as data classification, regression, denoising, dimensionality reduction, etc. In these contexts, deep learning algorithms have been successful in solving problems from different fields, including speech recognition, computer vision, language translation, etc. However, one criticism or drawback of deep neural networks is their lack of interpretability. In particular, traditional machine learning algorithms such as logistic regression, decision trees, etc. produce interpretable models that directly rank the relevance of features for a specific task. In certain problems such as anomaly detection in healthcare, it may be beneficial to explain the contribution of various factors to the prediction, in addition to achieving satisfactory performance analytically. Such interpretability is usually not possible with routine applications of artificial neural networks. Summary of the invention

[0005] In one embodiment, a feature ranking neural network is provided. According to the embodiment, the feature ranking neural network includes: an input layer, a broadcast layer encoding a plurality of weights, two or more separate and parallel branches of the feature ranking neural network, and an output layer downstream of the branches, wherein the output layer is configured to provide an output of the feature ranking neural network. The plurality of weights includes task-specific weights for each of a plurality of tasks for which the neural network is trained. The branches are configured to receive a set of filtered inputs from the broadcast layer based on the weights encoded by the broadcast layer. Each branch corresponds to a different task.

[0006] Furthermore, the weights of the broadcast layer filter the input features provided to the input layer to the above branches of the feature ranking neural network in a task-specific manner.

[0007] Further, the plurality of weights also includes common weights attributable to each task for which the neural network is trained.

[0008] Further, each weight in the plurality of weights is approximately zero or approximately one.

[0009] Further, each of the two or more branches of the feature ranking neural network includes a plurality of hidden layers.

[0010] Furthermore, multiple weights of the broadcast layer are learned by training a feature ranking neural network.

[0011] In another embodiment, a method for feature ranking is provided. According to the embodiment, a plurality of input samples are provided as a training data set to an input layer of a neural network. Each input sample is characterized by one or more features. The neural network includes a broadcast layer, the broadcast layer including a corresponding weight for each of a plurality of tasks. The input samples are processed to train the corresponding weight for each of the one or more features. For one or more of the plurality of tasks, based on the corresponding weight of the one or more features relative to the corresponding task, the one or more features are identified as being relevant to the corresponding task.

[0012] Furthermore, different tasks are characterized by different sets of recognition features.

[0013] Further, processing the input sample to train a corresponding weight for each task for each of the one or more features also includes training a common weight shared by the one or more features across multiple tasks.

[0014] Furthermore, the common weights are provided as part of the output of the feature ranking method.

[0015] Furthermore, the broadcast layer filters one or more features in a task-specific manner.

[0016] Further, the broadcast layer filters the one or more features to limit propagation of the one or more features to multiple parallel branches of the neural network downstream of the broadcast layer.

[0017] Furthermore, the method further includes: receiving a user input for further limiting the one or more features identified as being related to the corresponding task.

[0018] Furthermore, the method further includes: for each feature of the one or more features, applying a sparse constraint to the corresponding weight of each task.

[0019] Furthermore, the method further includes: applying similarity constraints to respective weights associated with two or more tasks having similar feature cardinalities based on user input.

[0020] Further, the one or more features include one or more biomarkers of interest, targeted therapies or treatments, electrocardiographic diagnostics, or neuroanalysis therapies or diagnostics.

[0021] In another embodiment, a method for generating a simplified feature set model is provided. According to the embodiment, one or more weights associated with a broadcast layer of a trained neural network are obtained. Each weight is associated with a corresponding feature and task combination. One or more task-specific features are identified for a given task based on the weights. A model is generated based on the one or more task-specific features, wherein the one or more task-specific features are subsets of a larger feature set for which the trained neural network is trained.

[0022] Furthermore, the model includes task-specific neural networks.

[0023] Further, in addition to the above-mentioned weights, the one or more task-specific features are identified based on user input.

[0024] Further, the method further includes using the model to evaluate additional observations relevant to the task based on the observed values ​​of the one or more task-specific features. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] These and other features, aspects and advantages of the present invention will be better understood when the following detailed description is read with reference to the accompanying drawings, in which like characters refer to like parts throughout the several views, and in which:

[0026] Figure 1 Depicted are examples of artificial neural networks for training deep learning models in accordance with aspects of the present disclosure;

[0027] Figure 2 A multi-task deep learning framework according to aspects of the present disclosure is described, wherein an input sample described by a feature can have multiple labels from K tasks;

[0028] Figure 3 Depicting a multi-task deep learning algorithm according to aspects of the present disclosure, wherein input features are processed by K neural networks for predicting the K tasks;

[0029] Figure 4 depicts an alternative representation of a multi-task deep learning algorithm according to aspects of the present disclosure, where input features are processed by K neural networks w for predicting the K tasks;

[0030] Figure 5 A chart depicting ranking features by importance for a non-limiting example implementation based on IMI detection according to aspects of the present disclosure;

[0031] Figure 6 is a block diagram of a computing device capable of implementing the method according to aspects of the present disclosure. DETAILED DESCRIPTION

[0032] One or more specific embodiments are described below. In order to concisely describe these embodiments, not all features of the actual implementation are described in the specification. It should be understood that 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 one implementation to another. In addition, it should be understood that such development work may be complex and time-consuming, but it is still a routine task of design, manufacturing and processing for ordinary technicians who benefit from this disclosure.

[0033] Although certain aspects of the following discussion are provided in the context of medical data analysis, it should be understood that the present technology is not limited to such a medical context. In fact, any examples and explanations provided herein in such a medical context are merely for the purpose of facilitating explanation by providing examples of certain real-world implementations and applications. However, the method of the present invention may also be used in other suitable data analysis contexts. In general, the method can be used in any data analysis context where interpretability of the output of an artificial neural network is required, for example to facilitate the selection of one or more features of interest in a relevant analysis context.

[0034] In view of the foregoing introductory comments, various aspects of the present method described herein utilize neural networks in a feature identification or ranking environment, and then the identified features may be selected for a specific application. As used herein, a feature is a measure for capturing or characterizing a phenomenon. Such a feature can be an individual measurable attribute or characteristic (such as activity, level, duration, constraint coefficient, etc.) of some observed phenomenon, although it is not necessarily an actual measurement and (although it) can take other forms such as qualitative characterization or observation. Neural networks as discussed herein can include deep neural networks, fully connected networks, convolutional neural networks (CNNs), perceptrons, autoencoders, regression networks, wavelet filter banks, or other neural network architectures. These techniques are referred to as deep learning techniques in this article, but the term can also be specifically used for the use of deep neural networks, which are neural networks with multiple layers.

[0035] As discussed herein, deep learning techniques (which may also be referred to as deep machine learning, hierarchical learning, or deep structured learning) are a branch of machine learning techniques that employ mathematical representations of data and artificial neural networks for learning. For example, deep learning methods are characterized in that they use one or more algorithms to extract or model a high-level abstraction of a class of data of interest. This may be accomplished using one or more processing layers, each of which typically corresponds to a different level of data abstraction and may therefore employ or utilize different aspects of the initial data or the output of a previous layer (i.e., layers of a hierarchy or cascade) as the goal of the process or algorithm at a given layer. In the context of data analysis, it is characterized by different layers corresponding to different levels of features or levels of abstraction in the data.

[0036] In general, processing from one level or abstraction to the next can be viewed as a "stage" of the analysis process. Each stage of the analysis can be performed by a separate neural network or by different parts of a larger neural network. For example, as discussed in this article, a single deep learning network can cover all stages in the analysis process (e.g., from initial input to output data set). Alternatively, each separate different deep learning network may cover only one stage (or a subset of the stages) of the overall analysis process.

[0037] As part of the initial training of a deep learning process to solve a particular problem, a training data set with known initial values ​​and known or expected values ​​of the final output of the deep learning process can be employed. A single stage of training can have known input values ​​corresponding to one representation space and known output values ​​corresponding to the representation space of the next level. In this way, the deep learning algorithm can process the known or training data set (in a supervised or guided manner or in an unsupervised or unguided manner) until a mathematical relationship between the initial data and the expected output is identified and / or a mathematical relationship between the input and output of each layer is identified and characterized. Similarly, a separate validation data set can be employed, where both the initial and expected target values ​​are known, but only the initial values ​​are provided to the trained deep learning algorithm, which is then compared to the output of the deep learning algorithm to verify previous training and / or prevent overtraining.

[0038] Considering the foregoing, Figure 1An example of an artificial neural network 50 is schematically depicted, which can be trained as a deep learning model as discussed herein. In this example, the network 50 is multi-layered, with a training input 52 and multiple layers including an input layer 54, hidden layers 58A, 58B, etc., as well as an output layer 60 and a training target 64 present in the network 50. In this example, each layer consists of a plurality of "neurons" or nodes 56. The number of neurons 56 can be constant between layers, or as shown, can vary between layers. The neurons 56 of each layer produce a corresponding output, which is used as the input of the neurons 56 of the next level. In practice, a weighted sum of the inputs with an additional bias is calculated to "stimulate" or "activate" each corresponding neuron of the layer according to a rectified linear unit (ReLU), a sigmoid function, a hyperbolic tangent function, or an activation function specified or programmed in other ways. The output of the last layer constitutes the network output 60, which is used together with a target value or construction 64 to calculate some loss or error function 62, which will be back-propagated to guide network training.

[0039] The loss or error function 62 measures the difference between the network output 60 and the training target 64. In some embodiments, the loss function can be the mean squared error (MSE). Alternatively, the loss function 62 can be defined by other metrics associated with the specific task in question, such as a softmax function.

[0040] In view of the foregoing, the neural network 50 can be trained for data analysis in this way: based on the relevance of relevant features to a given task so as to sort or identify relevant features, and possibly select some or all of these features for a given application or use. In particular, the present disclosure describes a method for performing feature sorting and / or selection within a deep neural network in a single task or multi-task / multi-label setting. A method for separating and discovering features specific to a single task or a limited subset of tasks (e.g., two or three tasks) and other features shared between multiple or all tasks is described. Applications of such methods include, but are not limited to, biomarker discovery and targeted therapy. In addition to obtaining interpretability, the identified feature subsets can be used to make leaner and / or more effective models, which are based only or primarily on those features identified as being particularly relevant to a given task, thereby resulting in better stability and regularization and reducing computational requirements and / or memory requirements. User input can also be used to identify the features most relevant to a given task (e.g., based on domain knowledge) and / or select from those features so identified for subsequent operations, such as forming a model based on a reduced or limited set of features.

[0041] Figure 2An example of a multi-task deep learning framework 100 is depicted, in which an input sample described by one or more features 102 may have multiple labels from K tasks corresponding to analysis results 108 (e.g., predictions). In one embodiment, the method discovers features shared among many or all of the K tasks 108 (shared features 110) and task-specific features (task-specific features 112) that are specific to a single task or a limited number of tasks (e.g., two or three tasks). The particularity of features in certain tasks may help reveal useful patterns or establish relationships that may enable better understanding of data and / or tasks and building intelligent systems. In the depicted example, multi-task feature ranking and / or selection is accomplished using one or more multi-task feature ranking neural networks or multiple parallel branches of a neural network (shown collectively using reference numeral 116) that are executed as a joint feature discovery engine for different tasks 108.

[0042] Go to Figure 3 and 4 , an example of a multi-task feature ranking network 116 is provided. In this example, the multi-task feature ranking network 116 takes the form of a multi-task deep learning algorithm, in which input features 102 (input via input layer 54) are processed by K neural networks or branches 120 of a neural network for predicting K tasks 122. To facilitate the implementation of the feature ranking method discussed herein, a broadcast layer 126 is provided that selects and / or filters the input features 102 before propagating through the separate and parallel branches 120 of the neural network. In one embodiment ( Figure 3 ), the broadcast layer 126 is an element-wise multiplication layer whose weights w are learned during neural network training. According to the present method, the weights w of the broadcast layer 126 reveal the relevance of the features 102 for different tasks 122, i.e., the larger the weight, the greater the relevance of the feature to a given task, which can facilitate the subsequent selection of one or more features relevant to a given task or application.

[0043] As an explanation, an input sample {X} is represented by N features 102. Let W be the set of weights of the feature weighting layer (i.e., broadcast layer 126). For a K-task learning problem, W has N*K weights, where for each task, N scaling factors W are learned. k .

[0044] One benefit of this approach is that it allows for joint discovery of task-specific features 112 (i.e., those features that contribute to a specific task or a limited number of tasks) and shared or common features 110 across multiple tasks 122 (i.e., those features that contribute to all or most tasks). The identified features ranked in this manner can then be selected from a given application or task. In one embodiment ( FIG. 4 ), the broadcast layer 126 weight W for each task is k is the shared task feature and the kth (k th ) is a linear addition of task-specific features, as shown in equation (1):

[0045] (1)

[0046] Among them, W s corresponds to the weight of the shared feature in task 122, and is the weight of the feature specific to task K.

[0047] To find a reduced set of relevant features 102, W k Imposing sparsity, as shown in Equation (2), one enforces the L1 norm to be D-sparse, where D ≤ N:

[0048] (2)|W k |≤D

[0049] Two approaches are contemplated for separating out shared features 110 in task 122 .

[0050] In the first method of expressing low-rank recovery, the weight W k are arranged as column vectors in W. From equations (1) and (2), it can be seen that the resulting matrix W should be a low-rank matrix. Shared feature weights W s can be recovered as a shared low-rank part of W. This can be achieved using traditional low-rank recovery techniques.

[0051] Alternatively, orthogonalization can be used as an explicit way to achieve feature separation. Learn (K+1)*N weights, where the extra N are used for shared features. Figure 4 As shown, using equation (1) to convert W s The role of is used for each task. Orthogonality is enforced between k pairs of weights: It is believed that shared features are not task-specific features and vice versa. Note that it is also possible to This can be done by including in the loss function (i.e. dot product) term to achieve this.

[0052] From the weight W k The absolute value of (n) infers the relevance (i.e., importance) of a particular feature n to task K, where:

[0053] (3)Importance(n,k)=abs(W k (n))

[0054] With this in mind, Table 1 shows the weight vector table after training:

[0055] Table 1

[0056] feature <![CDATA[W S ]]> <![CDATA[W i ]]> … <![CDATA[W j ]]> Feature 1 1.0 0.0 … 0.0 Feature 2 1.0 0.0 … 0.0 Feature 3 1.0 0.0 … 0.0 Feature 4 0.0 1.0 … 0.0 Feature 5 0.0 1.0 … 0.0 Feature 6 0.0 0.5 … 0.5 Feature 7 0.0 0.0 … 1.0 Feature 8 0.0 0.0 … 1.0 Feature 9 0.0 0.0 … 0.0 Feature 10 0.0 0.0 … 0.0 … … … … …

[0057] The trained weights may not be exactly 0 or 1 as is usually shown, but are expected to be close to 0 or 1. In some cases where low values ​​are observed, a value of 0 can be forced to increase the variance. Despite the sparsity constraint, some weight values ​​can be assumed to be close to 0.5, such as W j According to the weight, the feature can be roughly classified into the following categories based on the role it plays in multi-task network training: shared features (e.g., high W S Value, for example, W with a value of 1 S weights), task-specific features (e.g., high W i,j ,... values, such as weights with a value of 1), and / or redundant features (i.e., noise). With this classification, it is expected that each of these features is orthogonal to each other. The orthogonality constraint is added to the loss function using the sparsity constraint. In the example shown in Table 1, features 1-3 can be considered common or shared features because they have a common feature weight WS=1; features 4 and 5 are task-specific features for task i, with respective weights W i =1); Features 7 and 8 are task-specific features of task j, and their respective weights are W j =1); Features 9 and 10 are noise and have no non-zero weights for any specific task or are common to all tasks. In the case of feature 6, in this example, W i and W j The weights of are uniformly involved (e.g., a weight of 0.5), indicating that the feature is equally involved in both tasks. In this case, feature 6 is not a common or shared feature that contributes to all tasks, but is clearly (here equally) beneficial to only two (i.e., a limited number) of multiple tasks (here, tasks i and j), which can provide useful insights into the commonalities between these tasks.

[0058] Furthermore, based on domain knowledge, if it is known that certain groups of tasks have similar feature cardinality, this can be achieved by including (i.e., negative dot product) term to propose similarity constraints. Such domain knowledge and / or similarity constraints can be provided as user input to guide or direct the feature ranking process. Turning to Table 2, an example is provided, which shows a task with feature cardinality similarity, where for features 4 and 5, it can be seen that they have a similarity in weight W. x and W y The aspects are similar.

[0059] Table 2

[0060]

[0061]

[0062] In view of the foregoing, the present method can be used to explain deep neural networks without compromising the performance of the network. For problems where interpretability is necessary or beneficial, such as in radiology, anomaly detection, etc., these benefits can further illustrate the advantages of deep learning algorithms over classical machine learning algorithms. For example, in the context of biomarker discovery, the separation of shared features 110 and task-specific features 112 in a multi-task learning framework as discussed herein is essentially biomarker discovery and can facilitate prognosis, diagnosis, and treatment of diseases.

[0063] In addition, the present method can be used to build or improve the efficiency of lean models using subsets of features identified for each task, which will help reduce computational and memory usage. For example, based on the relevant features identified for a given task, a model or additional neural network can be generated based only on the relevant identified features. In some embodiments, the user can provide additional input in the process of selecting from the identified relevant features to further select which features to use in building a finite model or neural network. In other embodiments, all identified relevant features can be used in a finite model or neural network, or an automatic algorithm or routine can be used to select a subset of relevant features, such as automatically selecting a subset of relevant features based on maximum weight, based on reducing or limiting the covariance between selected features, etc. For example, a typical ECG interpretation includes predictions from approximately 1,000 different diagnostic codes. Using the current multi-task feature ranking network with a shared factor 'M' to build a neural network directly leads to a model that is leaner by 'M' times.

[0064] Further multi-task learning problems often suffer from problems such as class imbalance, sample diversity, etc., which can lead to unstable, unreliable networks. The reduction in parameters that can result from using feature ranking as discussed in this article can improve the stability of the neural network, where the broadcast layer 126 plays the role of learning regularization.

[0065] The following example provides a practical example illustration of the use and benefits of the present method. In this example, the feature ranking method discussed herein is applied to the problem of diagnosing the presence of inferior myocardial infarction (IMI) from electrocardiogram (ECG) data (i.e., a single task). Several measurements (interchangeably referred to as features in machine learning terms) are made on the raw ECG waveform used as input, while the presence or absence of IMI is considered as output. As discussed herein, the broadcast layer 126 discussed above is used to identify key features that contribute to prediction. These features are then associated with clinical features, which are referred to as contributing features to the presence of IMI.

[0066] The study dataset consists of 500,000 ECG cases with features extracted from 12 leads. Some of these features include: Q wave duration, Q wave amplitude, Q wave area, QR amplitude ratio, QRS axis, and T wave inversion.

[0067] A list of 72 features was used to predict the presence of IMI. A multilayer perceptron (MLP) neural network was trained to predict the presence of IMI. Three hidden layers 58 were used, each network consisting of 200 neurons each. Using this network, an accuracy of 96% was achieved. The sensitivity of the network was 79% and the specificity was 97%. According to the present method using a broadcast layer 126, a list of features and their importance or relevance scores for a given task were extracted (Equation (3)).

[0068] Based on this study, Figure 5 is a graph ranking features by their importance for IMI detection as defined by equation (3). The selected features were found to be associated with features known to be clinically important in IMI prediction. It should be noted that the diagnosis of IMI involves the detection of abnormal Q waves from a specific guide (e.g., aVF, II), and this is done by Figure 5 The corresponding Q-wave morphological features are captured by sorting.

[0069] In addition, additional experiments were performed using approximately half the number of initial features (35 features in this example) to further test the feature importance results. The resulting network provided similar results to the previous experiments using a limited number of features.

[0070] The proposed method is further tested on multiple tasks. Specifically, experiments are conducted on two important ECG-related diagnoses, namely anterior myocardial infarction (AMI) and inferior myocardial infarction (IMI). The dataset and features used are the same as in the previous experiments. The orthogonality constraint and the constraints in equations (1) and (2) are implemented as part of the feature ranking network. The features are classified according to equation (3) and three types of features are listed: 1) shared features, 2) IMI-specific features, and 3) AMI-specific features, as shown in Table 3:

[0071] Table 3

[0072]

[0073]

[0074] Strikingly, features related to Q wave morphology appeared as relevant features for both tasks, as these are Q wave abnormalities. Features identified as specific to IMI correspond to features observed in previous experiments. Features identified as specific to AMI correspond to previous guidance. Finally, shared features of Q Duration have been shown to be relevant, as described in the ECG literature. Therefore, features identified for both tasks appear to be clinically plausible.

[0075] Although the previous examples involve ECG, as mentioned above, the present method is not limited to such contexts, or to medical contexts in general. That is, the present method is general and not limited to ECG diagnostic or biomarker problems, but can also be applied to problems such as neuroanalysis or non-healthcare applications.

[0076] It should be understood that one can use Figure 6 A processor-based system as shown in or several such systems in communication with each other may be used to perform or implement some or all of the methods discussed herein related to feature ranking using a trained deep neural network. Such a system may include Figure 6 Some or all of the computer components depicted in . Figure 6 A block diagram generally illustrates example components of a computing device 150 and their potential interconnections or communication paths, such as along one or more buses. As used herein, computing device 150 may be implemented as one or more computing systems, including laptops, notebooks, desktops, tablets, or workstations, as well as server-type devices or portable communication-type devices and / or other suitable computing devices.

[0077] As shown, computing device 150 may include various hardware components such as one or more processors 152, one or more buses 154, memory 156, input structures 158, power supply 160, network interface 162, user interface 164, and / or other computer components for performing the functions described herein.

[0078] In some embodiments, the one or more processors 152 are microprocessors configured to execute instructions stored in the memory 156 or other accessible location. Alternatively, the one or more processors 152 may be implemented as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), and / or other devices designed to perform the functions discussed herein in a dedicated manner. It should be appreciated that multiple processors 152 or processing components may be used to perform the functions discussed herein in a distributed or parallel manner.

[0079] Memory 156 may include any tangible, non-transitory medium for storing data or executable routines, including volatile memory, non-volatile memory, or any combination thereof. Although shown for convenience as Figure 6 Although the memory 156 is a single block in a memory 156, it can actually contain various discrete media in the same or different physical locations. One or more processors 152 can access data in the memory 156 via one or more buses 154.

[0080] Input structures 158 are used to allow a user to input data and / or commands to the device 150, and may include a mouse, a touchpad, a touch screen, a keyboard, etc. Power supply 160 may be any suitable energy source for providing power to the various components of the computing device 150, including line and battery power. In the depicted example, the device 100 includes a network interface 162. Such a network interface 162 may allow communication with other devices on a network using one or more communication protocols. In the depicted example, the device 100 includes a user interface 114, such as a display configured to display an image or date provided by one or more processors 152.

[0081] The technical effects of the present invention include ranking features using deep neural networks in multi-task and / or multi-label settings. Task-specific features as well as features shared between multiple tasks can be identified. In addition to promoting interpretability, the selected feature subsets can also be used to make effective models, thereby achieving better stability and regularization while reducing computation and memory usage, thereby achieving more efficient implementation.

[0082] This written description uses examples to disclose the invention, including the best mode, and also to enable those skilled in the art to practice the invention, including making and using any device or system and performing any combined methods. The patentable scope of the invention is defined by the claims, and may include other examples that occur to those skilled in the art. These other examples are intended to be included within the scope of the claims if they have structural elements that do not differ from the literal language of the claims, or if they include equivalent structural elements that are not substantially different from the literal language of the claims.

Claims

1. A computing device, comprising: processor; as well as Feature ranking neural network, including: Input layer, used to input ECG data; a broadcast layer for encoding a plurality of weights, wherein the plurality of weights comprises a task-specific weight for each of a plurality of tasks for which the neural network was trained, the broadcast layer for identifying features of the electrocardiogram data; two or more separate and parallel branches of the feature ranking neural network, wherein the branches are configured to receive a set of filtered inputs from the broadcast layer based on weights encoded by the broadcast layer, wherein each branch corresponds to a different task, wherein the set of filtered inputs includes features of the electrocardiogram data; and An output layer downstream of the branch, wherein the output layer is configured to provide an output of the feature ranking neural network, wherein the output comprises the presence of inferior wall myocardial infarction.

2. The computing device of claim 1, wherein: The weights of the broadcast layer filter the input features provided to the input layer to the branch of the feature ranking neural network in a task-specific manner.

3. The computing device of claim 1, wherein: The plurality of weights also includes a common weight attributable to each task for which the neural network is trained.

4. The computing device of claim 1, wherein: Each weight of the plurality of weights is approximately zero or approximately one.

5. The computing device of claim 1, wherein: Each of the two or more branches of the feature ranking neural network includes a plurality of hidden layers.

6. The computing device of claim 1, wherein: The plurality of weights of the broadcast layer are learned by training the feature ranking neural network.

7. A feature ranking method, comprising: providing a plurality of input samples as a training data set to an input layer of a neural network, wherein each input sample is characterized by one or more features, wherein the input samples include electrocardiogram data; processing the input samples at a broadcast layer of the neural network to train a corresponding weight for each task for each of the one or more features, wherein the broadcast layer includes a corresponding weight for each of a plurality of tasks, the broadcast layer being used to identify features of the electrocardiogram data; For one or more tasks in the plurality of tasks, identifying the one or more features as being relevant to the respective task based on respective weights of the one or more features relative to the respective task; The existence of inferior wall myocardial infarction is outputted at the output layer of the neural network.

8. The method according to claim 7, characterized in that Different tasks are characterized by different sets of recognition features.

9. The method according to claim 7, characterized in that Processing the input samples to train a corresponding weight for each task for each of the one or more features also includes training a common weight shared by the one or more features across the plurality of tasks.

10. The method according to claim 9, characterized in that The common weights are provided as part of the output of the feature ranking method.

11. The method according to claim 7, characterized in that The broadcast layer filters the one or more features in a task-specific manner.

12. The method according to claim 7, characterized in that The broadcast layer filters the one or more features to limit propagation of the one or more features to multiple parallel branches of the neural network downstream of the broadcast layer.

13. The method according to claim 7, characterized in that Also includes: User input is received to further restrict the one or more features identified as being relevant to the respective task.

14. The method according to claim 7, characterized in that Also includes: For each of the one or more features, a sparse constraint is imposed on a corresponding weight of each task.

15. The method according to claim 7, characterized in that Also includes: Based on user input, a similarity constraint is imposed on respective weights associated with two or more tasks having similar feature cardinality.

16. The method according to claim 7, characterized in that The one or more characteristics include an electrocardiogram diagnosis.

17. A method for generating a simplified feature set model, comprising: Enter ECG data; Obtaining one or more weights associated with a broadcast layer of a trained neural network, wherein each weight is associated with a corresponding feature and task combination, wherein the broadcast layer is used to identify features of the electrocardiogram data; identifying one or more task-specific features for a given task based on the electrocardiogram data and the weights; and A model is generated based on the one or more task-specific features, the model comprising a task-specific neural network, wherein the one or more task-specific features are a subset of a larger feature set on which the trained neural network was trained, wherein an output of the model comprises the presence of inferior wall myocardial infarction.

18. The method according to claim 17, characterized in that In addition to the weights, the one or more task-specific features are identified based on user input.

19. The method according to claim 17, characterized in that Also included is using the model to evaluate additional observations relevant to the task based on the observed values ​​of the one or more task-specific features.

Images