automatic sleep staging using state machine-controlled decision trees

18
Automatic Sleep Staging Using State Machine-controlled Decision Trees Syed Anas Imtiaz Esther Rodriguez-Villegas

Upload: anas-imtiaz

Post on 23-Jan-2018

293 views

Category:

Engineering


3 download

TRANSCRIPT

Page 1: Automatic Sleep Staging Using State Machine-controlled Decision Trees

Automatic Sleep Staging Using State Machine-controlled Decision Trees

Syed Anas Imtiaz Esther Rodriguez-Villegas

Page 2: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

2

INTRODUCTION

Sleep staging has been an active research area for over four decades

Involves analysis of physiological signals and classification into one of Wake, N1, N2, N3 or REM stages

Typically uses ExG (EEG/EOG/EMG) signals

Actigraphy, HRV and respiratory increasingly used now

Recent consumer focus with sleep analysis integrated with wearables and fitness trackers

Page 3: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

3

…for sleep tracking

Wearable Devices

Actigraphy, accelerometry, heart rate

Difficult to distinguish between all stages

Over/under estimation of sleep time

Medical usage still questionable

May improve with time

Not as accurate as EEG

Page 4: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

4

SLEEP STAGING RESEARCH

0

5

10

15

20

1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015

Number of papers indexed in IEEEXplore over the last 25 years

Page 5: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

5

SLEEP STAGING RESEARCH

Most algorithms have very good classification accuracy

However, this performance comes at a cost

COMPUTATIONAL COMPLEXITY: ANN, SVM, Large number of features, etc.

On hardware this translates to higher POWER CONSUMPTION

Limited power budget in wearable systems

Page 6: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

6

OUR AIM

To develop an automatic sleep staging algorithm that is suitable for use in resource-constrained wearable systems “

Extract spectral features and classify using a decision tree (DT)

DTs are amongst the simplest of classifiers

Use small contextually-aware DTs rather than one long tree

Order of execution controlled by a state machine

Page 7: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

7

DATABASE

PhysioNet Sleep EDF Expanded Database

61 PSG recordings

Almost 60,000 epochs of 30-second duration

31 recordings in training set and 30 in test set

Two EEG channels: Fpz-Cz and Pz-Oz

Page 8: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

8

FEATURES

Channel Features

Fpz-Cz

sigma/beta, beta/delta, delta/alpha, beta/alpha, SEFd(8-16 Hz), SEFd(0.5-8 Hz), SEF95(0.5-30 Hz), SEF50(0.5-8 Hz), line length (11-16 Hz), rel. delta2, rel. beta, rel. gamma, abs. delta, abs. delta1, abs. delta2, abs. alpha2

Pz-Oz sigma/beta, beta/delta, theta/alpha, beta/alpha, SEF95(0.5-30 Hz), SEF50(0.5-8 Hz), rel. beta, rel. gamma, rel. alpha, rel. theta, abs. delta, abs. delta1, abs. alpha1

66 features initially

Each feature for an epoch is an average for fifteen 2-second subepochs

Top 30 features using sequential feature selection (SFS)

Page 9: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

9

DESIGN OF DECISION TREES

Decision trees, in comparison to more complex classifiers, are only beneficial when the number of nodes in tree is relatively small

Use features with a decision tree

But limiting the number of nodes can directly affect the performance

Trade-offs needed

Alternative approach?

Page 10: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

10

OUR APPROACH

When a certain sleep stage is prevalent, it normally stays for a few epochs before transitioning to the next stage. “

An epoch only needs to be tested to check whether it is of the same sleep stage or not

One-vs-all decision tree

If transition needed, use a series of one-versus-one decision trees

Order of execution based on the likelihood of next sleep stage

Combination of decision trees contextually driven by a state machine

Activated based on the current sleep stage

Approach commonly used in AI for game development

Page 11: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

11

HOW IT WORKS?

Initial State

Check if the new epoch is also Wake or not

If Wake, maintain state and return

Assign Wake if all tests fail

Test for other sleep stages and exit when different stage found

Page 12: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

12

HOW IT WORKS?

New State

Check if the new epoch is also N2 or not

If N2, maintain state and return

Assign N2 if all tests fail

Test for other sleep stages and exit when different stage found

Page 13: Automatic Sleep Staging Using State Machine-controlled Decision Trees

13

RESULTS

Page 14: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

14

Using 30 PSG recordings for PhysioNet EDF Expanded Database

TRAINING RESULTS

WAKE 84.3%

N1 29.8%

N2 88.5%

N3 81.8%

Of the 29499 epochs in this set, 24255 are correctly classified

resulting in an overall accuracy of 82.22%

All stages show sensitivity of over

80% except N1

REM 87.4%

Page 15: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

15

Using 31 PSG recordings for PhysioNet EDF Expanded Database

TEST RESULTS

WAKE 72.1%

N1 22.0%

N2 88.2%

N3 79.6%

Of the 29817epochs in this set, 23512 are correctly classified

resulting in an overall accuracy of 78.85%

Similar results in N2, N3 and REM

and a noticeable reduction in sensitivity for Wake an N1

REM 84.1%

Page 16: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

16

DISCUSSION

About 80% classification performed by the Core decision trees.

Average number of comparisons for a decision is 3.4 (longest 12, shortest 2).

Total number of comparisons reduced by over 40% compared to a traditional decision tree for similar classification performance.

Only a subset of features and trees needed in each state resulting in better usage of processor resources consequently reducing power consumption.

Page 17: Automatic Sleep Staging Using State Machine-controlled Decision Trees

IEEE Engineering in Medicine and Biology Conference 2015

17

FUTURE WORK

Search for more characteristic features with better discriminatory ability and potentially reduce number of features.

BETTER FEATURES

Implement the algorithm on low-power microcontroller and/or integrated circuit to measure actual power consumption.

HARDWARE IMPLEMENTATION

The peripheral decision trees to discriminate between two stages can be designed with different weights assigned.

IMPROVED DESIGN OF TREES

Check to see if algorithm is useful with recordings from others sources and ways of improving performance.

MORE TESTING

Page 18: Automatic Sleep Staging Using State Machine-controlled Decision Trees

Q u e s t i o n s / C o m m e n t s

Thank You