Energy-efficient on-chip classifier for detecting physiological conditions

Patent metadata

Field	Value
Publication number	US-11682492-B2
Application number	US-202016946151-A
Country	US
Kind code	B2
Filing date	Jun 8, 2020
Priority date	Jun 7, 2019
Publication date	Jun 20, 2023
Grant date	Jun 20, 2023

Abstract

Methods, systems, and devices are disclosed for an efficient hardware architecture to implement gradient boosted trees for detecting biological conditions. For example, a method of detecting a biological condition includes receiving, by a device, a plurality of physiological signals from a plurality of input channels of the device, selecting, based on a trained prediction model, one or more input channels from the plurality of input channels, converting the one or more physiological signals received from the one or more input channels to one or more digital physiological signals, identifying, by using the plurality of gradient boosted decision trees, the selected characteristic in the one or more digital physiological signals, and determining a presence of a physiological condition based on an addition of the output values obtained from the plurality of gradient boosted decision trees.

First claim

What is claimed is:

1. A method of detecting a biological condition, comprising:
receiving, by a multiplexer of an implantable sensor device, a plurality of physiological signals from a plurality of input channels of the device;
selecting, based on a trained prediction model, one or more input channels from the plurality of input channels received by the multiplexer, wherein the trained prediction model indicates the one or more input channels and configurations of a plurality of gradient boosted decision trees for identification of a selected feature of one or more physiological signals from the plurality of physiological signals, wherein each of the trees has a plurality of nodes and each of the nodes is associated with a respective one of the channels in memory;
converting the one or more physiological signals received from the one or more input channels selected by the multiplexer to one or more digital physiological signals;
selecting coefficients for a programmable finite impulse response (FIR) filter of a feature extraction engine (FEE) for the selected one or more input channels from a memory based on a feature to be processed by the programmable FIR filter;
inputting the one or more digital physiological signals to the FEE, wherein the FIR filter extracts a feature from the one or more digital physiological signals, the FEE being connected to a comparator, the comparator receiving an input feature from the FEE and implementing one of the plurality of gradient boosted decision trees, wherein the comparator compares the input feature to a threshold value for a current node of the one of the plurality of gradient boosted decision trees and generates an output of the comparator for the current node;
determining for the current node, using the output of the comparator, whether a next node is a leaf node, and:
if the next node is not a leaf node, returning to selecting of one or more input channels;
if the next node is a leaf node, output a result value; and

determining a presence of a physiological condition based on an addition of the result values obtained from the plurality of gradient boosted decision trees.

2. The method of claim 1 ,
wherein a plurality of gradient boosted decision trees operate in parallel,
wherein the generation of the output of the comparator is performed within an optimum time that is determined based on a plurality of times associated with the plurality of gradient boosted decision trees, and
wherein each of the plurality of times indicate an amount of time associated with obtaining a result value from an associated gradient boosted decision tree.

3. The method of claim 2 ,
wherein a number of the plurality of gradient boosted decision trees is up to eight, and
wherein each gradient boosted decision tree has a maximum pre-determined depth of four.

4. The method of claim 1 ,
wherein the programmable FIR filter includes a first stage that outputs a downsampled physiological signal that is obtained by downsampling the digital physiological signal,
wherein the programmable FIR filter includes a second stage that includes a tunable bandpass filter that filters the downsampled physiological signal, and
wherein bandwidth related parameters of the tunable bandpass filter are determined based on the selected feature.

5. The method of claim 4 , wherein any one or more of the first stage and the second stage are bypassed based on the selected feature.

6. The method of claim 1 , wherein the selecting, the converting, and the determining is performed for the one or more input channels that are selected without buffering data from the plurality of input channels other than the one or more input channels.

7. An implantable sensor device receiving a plurality of physiological signals from a plurality of input channels, comprising;
a multiplexer receiving the plurality of physiological signals from the plurality of input channels;
a classifier for identifying a selected feature of one or more physiological signals from the plurality of physiological signals received by the multiplexer based on a trained prediction model that indicates the one or more input channels and configurations of a plurality of gradient boosted decision trees, each decision tree of the plurality of gradient boosted decision trees comprising a single feature extraction engine (FEE) that integrates a programmable finite impulse response (FIR) filter for multi-band filtering and spectral feature extraction, a digital comparator, and a Tree Control Unit (TCU), the FEE being connected to the comparator receiving an input from the FEE;
a memory configured to store the trained prediction model having coefficient values based on the selected features for the programmable FIR filter and feature thresholds for each gradient boosted decision tree;
a processor configured to send control signals to each FEE based on the trained prediction model for a respective gradient boosted decision tree to:
select an input channel from the plurality of input channels received by the multiplexer, the input channel associated with a node of the respective gradient boosted decision tree;
select coefficients for the programmable FIR filter from the memory based on a feature to be processed by the programmable FIR filter for the selected input channel wherein the FIR filter extracts a feature from the physiological signals received by the input channel, the comparator receiving the input from the FEE and implementing one of the plurality of gradient boosted decision trees, wherein the comparator compares the input to a threshold value for a current node of the one of the plurality of gradient boosted decision trees and generates an output of the comparator for the current node;
determine for the current node, using the output of the comparator whether a next node is a leaf node, and:
if the next node is not a leaf node, return to of selecting an input channel from the plurality of input channels;
if the next node is a leaf node, output a result value;

determine a presence of a physiological condition based on an addition of the result values obtained from the plurality of gradient boosted decision trees.

8. The device of claim 7 ,
wherein the processor is configured to identify the feature within an optimum time that is determined based on a plurality of times associated with the plurality of gradient boosted decision trees, and
wherein each of the plurality of times indicate an amount of time associated with obtaining a result value from an associated gradient boosted decision tree.

9. The device of claim 7 ,
wherein the programmable FIR filter includes a first stage that is configured to output a downsampled physiological signal obtained by downsampling the digital physiological signal,
wherein the programmable FIR filter includes a second stage that includes a tunable bandpass filter that is configured to filter the downsampled physiological signal, and
wherein bandwidth related parameters of the tunable bandpass filter are determined based on the selected characteristic.

10. The device of claim 9 , wherein any one or more of the first stage and the second stage are bypassed based on the selected characteristic.

11. A non-transitory machine-readable medium having machine executable instructions stored thereon that, when executed by one or more processors, direct the one or more processors to perform a method comprising:
receiving, by a multiplexer of an implantable sensor device, a plurality of physiological signals from a plurality of input channels of the device;
selecting, based on a trained prediction model, one or more input channels from the plurality of input channels received by the mul

Assignees

• Univ Cornell
• California Inst Of Techn

Inventors

• Shoaran Mahsa
• Taghavi Milad
• Haghi Benyamin
• Farivar Masoud
• Emami Azita

Classifications

• G16H50/30

  for calculating health indices; for individual health risk assessment · CPC title

• G16H40/67

  for remote operation · CPC title

• A61B5/1101

  Detecting tremor · CPC title

• A61B5/4094

  Diagnosing or monitoring seizure diseases, e.g. epilepsy · CPC title

• G16H50/70

  for mining of medical data, e.g. analysing previous cases of other patients · CPC title