Systems and methods for using federated learning for training centralized seizure detection and prediction models on decentralized datasets

Patent metadata

Field	Value
Publication number	US-2021407678-A1
Application number	US-202117356342-A
Country	US
Kind code	A1
Filing date	Jun 23, 2021
Priority date	Jun 24, 2020
Publication date	Dec 30, 2021
Grant date	—

Abstract

A server for updating a current version of a machine learning model resident in implanted medical devices includes an interface, a memory, and a processor. The interface is configured to receive a plurality of updated versions of the machine learning model from a plurality of remote sources remote from the server. The remote source may be, e.g., implanted medical devices and/or subservers. The processor is coupled to the memory and the interface and is configured to aggregate the plurality of updated versions to derive a server-updated version of the machine learning model, and to transmit the server-updated version of the machine learning model to one or more of the plurality of remote sources as a replacement for the current version of the machine learning model.

First claim

What is claimed is:

1 . A method of updating a current version of a machine learning model resident in a plurality of implanted medical devices, the method comprising:
receiving, at a server, a plurality of updated versions of the machine learning model from a plurality of remote sources remote from the server; aggregating, at the server, the plurality of updated versions to derive a server-updated version of the machine learning model; and transmitting, at the server, the server-updated version of the machine learning model to one or more of the plurality of remote sources as a replacement for the current version of the machine learning model.

2 . The method of claim 1 , wherein the machine learning model is configured to detect a neurological event, to predict an occurrence of a neurological event, or to initiate a delivery of a stimulation therapy.

3 . The method of claim 1 , wherein:
the machine learning model comprises a neural network architecture having a plurality of nodes, and is characterized by a plurality of biases, each of the plurality of biases being associated with a corresponding node of the plurality of nodes, and aggregating the plurality of updated versions comprises:
for at least one node of the plurality of nodes included in the plurality of updated versions of the machine learning model, calculating an average of the biases associated with the at least one node, and
assigning the average to the at least one node.

4 . The method of claim 3 , wherein aggregating the plurality of updated versions further comprises:
prior to calculating an average of the biases, applying a weight factor to each of the biases associated with the at least one node, wherein each weight factor is based on an amount of data included in a dataset on which the updated version of the machine learning model was trained.

5 . The method of claim 1 , wherein:
the machine learning model comprises a neural network architecture having a plurality of nodes and a plurality of interconnections between pair of nodes of the plurality of nodes, and is characterized by a plurality of weights, each weight of the plurality of weights being associated with a corresponding one of the plurality of interconnections, and aggregating the plurality of updated versions comprises:
for at least one connection of the plurality of interconnection included in the plurality of updated versions of the machine learning model, calculating an average of the weights associated with the at least one interconnection, and
assigning the average to the at least one interconnection.

6 . The method of claim 5 , wherein aggregating the plurality of updated versions further comprises:
prior to calculating an average of the weights, applying a weight factor to each of the weights associated with the at least one interconnection, wherein each weight factor is based on an amount of data included in a dataset on which the updated version of the machine learning model was trained.

7 . The method of claim 1 , wherein:
the machine learning model comprises a neural network architecture having a plurality of nodes, and is characterized by a plurality of biases, each of the plurality of biases being associated with a corresponding node of the plurality of nodes, and aggregating the plurality of updated versions comprises:
grouping the plurality of nodes into one or more sets of nodes based on one of probabilistic federated neural matching or federated matched averaging;
for at least one of the sets of nodes, calculating an average of the biases associated with the nodes in the at least one set of nodes, and
assigning the average to the nodes included in the at least one set of nodes.

8 . The method of claim 1 , wherein the machine learning model comprises one of: a logistic regression having one or more parameters; a convolutional neural network (CNN); an autoencoder; and a recurrent neural network (RNN).

9 . The method of claim 1 , wherein the plurality of remote sources comprises one or more of a plurality of implanted medical devices (IMD), and further comprising:
generating, at the one or more of the implanted medical devices, an IMD-updated version of the machine learning model based on the current version of the machine learning model and a dataset stored in the implanted medical device.

10 . The method of claim 9 , wherein generating an IMD-updated version of the machine learning model comprises:
extracting, at the one or more of the implanted medical devices, features from a plurality of physiological records included in the dataset; and training the machine learning model on the extracted features.

11 . The method of claim 10 , wherein each of the plurality of physiological records is of a same type, comprising any one of:
electrical activity of a brain, neural tissue motion, heart rate, blood profusion, blood oxygenation, neurotransmitter concentrations, blood glucose, sweat hormones, body motion, and pH level.

12 . The method of claim 10 , wherein each of the plurality of physiological records has a same tag, which identifies a common aspect among the plurality of physiological records, the common aspect corresponding to one of:
an occurrence of a neurological event; absence of a neurological event; and patient state.

13 . The method of claim 9 , further comprising transmitting, at the one or more of the implanted medical devices, the IMD-updated version of the machine learning model to the server, wherein the IMD-updated version corresponds to one of the plurality of updated versions aggregated at the server.

14 . The method of claim 9 , further comprising:
transmitting, at the one or more of the implanted medical devices, the IMD-updated version of the machine learning model to a subserver remote from the server; aggregating, at the subserver, the IMD-updated versions to derive a subserver-updated version of the machine learning model; and transmitting, at the one or more subservers, the subserver-updated version of the machine learning model to the server, wherein the subserver-updated version corresponds to one of the plurality of updated versions aggregated at the server.

15 . The method of claim 1 , wherein the plurality of remote sources comprises one or more subservers remote from the server, and further comprising:
generating, at the one or more subservers, a subserver-updated version of the machine learning model based on a dataset received by the subserver from one or more of the plurality of implanted medical devices; and transmitting, at the one or more subservers, the subserver-updated version of the machine learning model to the server, wherein the subserver-updated version corresponds to one of the plurality of updated versions aggregated at the server.

16 . The method of claim 15 , wherein generating a subserver-updated version comprises:
pooling, at the one or more subservers, a plurality of datasets received from the one or more of the plurality of implanted medical devices to create a dataset pool; and training, at the one or more subservers, the current version of the machine learning model on the dataset pool.

17 . The method of claim 16 , wherein training comprises:
extracting, at the one or more subservers, features from a plurality of physiological records; and training the machine learning model on the extracted features.

18 . The method of claim 15 , wherein generating a subserver-updated version comprises:
for one or more of the implanted medical devices from which the one or more subs

Assignees

• Neuropace Inc

Inventors

• Arcot Desai Sharanya
• Tcheng Thomas K

Classifications

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/098 Primary

  Distributed learning, e.g. federated learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title