EP-4738146-A1 - NEURAL NETWORK FOR VARIABLE NUMBER OF SIGNALS

Abstract

Proposed concepts thus aim to provide schemes, solutions, concepts, designs, methods and systems pertaining to processing a variable number of clinical measurement signals. In particular, embodiments aim to provide a neural network for processing a variable number of clinical measurement signals by including a plurality of selectable input layers such that the neural network can adapt to the number of clinical measurement signals being input into it.

Inventors

• BRESCH, Erik
• GROSSEKATHOEFER, Ulf

Assignees

• Koninklijke Philips N.V.

Dates

Publication Date

20260506

Application Date

20241101

Claims (15)

1. A neural network (100) for processing a variable number of clinical measurement signals, the neural network comprising: a plurality of selectable input layers (110), each selectable input layer configured to, in use, receive an input clinical measurement signal, wherein each clinical measurement signal comprises respective clinical data of a same subject, and wherein each selectable input layer shares a first output tensor size; at least one selectable trunk (120) having a first input tensor size, equal to the first output tensor size of the selectable input layers, and a second output tensor size; and at least one output layer (130) configured to, in use, output at least one predicted feature of the one or more input clinical measurement signals; wherein the neural network is configured to, in use: select a number of the plurality of selectable input layers equal to the number of input clinical measurement signals.
2. The neural network of claim 1, further comprising a plurality of selectable trunks (220), wherein each selectable trunk shares the first input tensor size and the second output tensor size; and wherein the neural network is configured to, in use: select at least one of the plurality of selectable trunks based on available computational power of a processing arrangement running the neural network.
3. The neural network of claim 1 or 2, further comprising a plurality of streaming convolutional layers (221).
4. The neural network of claim 3, wherein the plurality of streaming convolutional layers are in at least one of the plurality of selectable trunks.
5. The neural network of claim 3 or 4, further comprising, downstream of the plurality of streaming convolutional layers, at least one of: a long-short-term memory unit (225); and a general recursive unit (225).
6. The neural network of any of claims 1 to 5, further comprising a plurality of selectable output layers (230), each output layer configured to, in use, output a different predicted feature of the one or more input clinical measurement signals.
7. The neural network of any of claims 1 to 6, wherein the at least one predicted feature comprises at least one of: wave delineation; detected beats; beat classification; an alert; ST segment length; rhythm class; and a modified version of at least one of the input clinical measurement signals.
8. The neural network of any of claims 1 to 7, wherein weights of the neural network are quantized and/or compressed.
9. The neural network of any of claims 1 to 8, wherein the one or more input clinical measurement signals comprise at least one of: one or more input ECG lead signals; one or more input Pleth signals: and one or more input invasive blood pressure signals.
10. A computer-implemented method (400) for processing a variable number of clinical measurement signals, the method comprising: providing input data (410) comprising one or more clinical measurement signals, wherein each clinical measurement signal comprises respective clinical data of a same subject, to a neural network, the neural network being trained to predict, from the input data, at least one feature of the input data; wherein the neural network comprises: a plurality of selectable input layers, wherein each selectable input layer shares a first output tensor size; at least one selectable trunk having a first input tensor size, equal to the first output tensor size of the selectable input layers, and a second output tensor size; and at least one output layer for outputting the at least one predicted feature of the input data; selecting a number of the plurality of selectable input layers (420) equal to the number of clinical measurement signals.
11. A computer-implemented method (500) for training a neural network for processing a variable number of clinical measurement signals, the neural network comprising a plurality of selectable input layers, wherein each selectable input layer shares a first output tensor size; at least one selectable trunk having a first input tensor size, equal to the first output tensor size of the selectable input layers, and a second output tensor size; and at least one output layer for outputting at least one predicted feature of input data, wherein the method comprises: training the neural network (510) on training data comprising a plurality of different numbers of clinical measurement signals
12. The computer-implemented method of claim 11, wherein training the neural network comprises training the neural network with augmentation.
13. The computer-implemented method of claim 12, wherein the augmentation comprises at least one of: clinical measurement signal dropout; clinical measurement signal swap; gain change; offset change; time-shift; and noise injection.
14. The computer-implemented method of any of claims 11 to 13, the neural network comprises n selectable input layers, and wherein the training data comprises every permutation of the presence or absence of n clinical measurement signals.
15. A computer program comprising code means for implementing the method of any of claims 10 to 14 when said program is run on a processing system.

Description

FIELD OF THE INVENTION This invention relates to the field of neural networks, and in particular to the field of neural networks for processing clinical measurement signals. BACKGROUND OF THE INVENTION In a clinical environment, clinical measurement data, such as ECG data, typically consists of one or more lead signals. The number of clinical measurement signals present/available in any one situation, however, is variable. Furthermore, neural networks, typically have a fixed input tensor size, which makes it difficult to process multi-dimensional signals with varying dimensionality when deployed (e.g., 2-lead ECG vs 12-lead ECG). Constructing a neural network for the maximum number of possible inputs and then zero-filling unused inputs is not a desirable solution, however, because the network then cannot intrinsically understand the difference between, for example, a zero signal because the input is unused, a zero signal because, for example, an ECG electrode fell off, and a zero signal because the subject has died. Furthermore, the coefficients of neural networks can take up a sizable amount of storage space which is a valuable resource in embedded systems. Therefore, simultaneously deploying several networks (i.e., one for each possible input lead configuration) is a highly undesirable solution. SUMMARY OF THE INVENTION The invention is defined by the claims. According to examples in accordance with an aspect of the invention, there is provided a neural network for processing a variable number of clinical measurement signals. The neural network comprises: a plurality of selectable input layers, each selectable input layer configured to, in use, receive an input clinical measurement signal, wherein each clinical measurement signal comprises respective clinical data of a same subject, and wherein each selectable input layer shares a first output tensor size; at least one selectable trunk having a first input tensor size, equal to the first output tensor size of the selectable input layers, and a second output tensor size; and at least one output layer configured to, in use, output at least one predicted feature of the one or more input clinical measurement signals; wherein the neural network is configured to, in use: select a number of the plurality of selectable input layers equal to the number of input clinical measurement signals. Proposed concepts thus aim to provide schemes, solutions, concepts, designs, methods and systems pertaining to processing a variable number of clinical measurement signals. In particular, embodiments aim to provide a neural network for processing a variable number of clinical measurement signals by including a plurality of selectable input layers such that the neural network can adapt to the number of clinical measurement signals being input into it. It is proposed that by providing a neural network structure which has multiple input layers which can be selected or not, the neural network can adapt to however many clinical measurement signals are input into it, and therefore be able to distinguish between situations where a signal is not present rather than the signal containing a value of zero (or just noise). This provides a highly efficient way of dealing with a variable number of clinical measurement signals, not requiring multiple entirely separate neural networks. By ensuring that each input layer has the same output tensor size, which is equal to the input tensor sizes of each trunk present in the neural network, the neural network can effectively `mix and match' the different layers and trunks. Furthermore, by specifying that the number of input layers is selected to be equal to the number of input clinical measurement signals, the neural network remains as efficient as possible (in this respect) while also preventing situations where an input zero-signal is simply because, for example, an ECG lead is not being used, which would confuse the output of the neural network. In addition, facilitating the provision of one or more selectable output layers allows each output layer to be specifically trained to output a specific predicted feature (or set of predicted features) of the clinical measurement signals. Embodiments may be of particular use with ECG lead signals, as the number of ECG leads used to monitor a patient often varies depending on the context. Ultimately, an improved neural network for processing a variable number of clinical measurement signals may be provided. In some embodiments, the neural network may further comprise a plurality of selectable trunks, wherein each selectable trunk shares the first input tensor size and the second output tensor size; and wherein the neural network may be configured to, in use: select at least one of the plurality of selectable trunks based on available computational power of a processing arrangement running the neural network. This may provide a neural network structure which can adapt to the computational power of the proc