1 of 20

ESTIMATION OF RESPIRATORY RATE FROM PPG SIGNALS USING DEEP NEURAL NETWORK

Submitted by,

Mst. Shamima Hossain

Student ID-1406128

Submitted to,

Dr. Md. Kamrul Hasan

Professor, Dept. of EEE,

BUET

2 of 20

Presentation Outline

Overview

Current Methods

Proposed Model

Result Analysis

Future Work

3 of 20

Thesis Overview

*bpm=breaths per minute

What is Respiratory Rate (RR)?

number of breaths you take per minute
normal RR for an adult at rest is 12 to 20 bpm*

Why Using PPG?

Automatic detection

Causes slow variation in DC component

4 of 20

Thesis Overview

*bpm=breaths per minute

What is Respiratory Rate (RR)?

number of breaths you take per minute
normal RR for an adult at rest is 12 to 20 bpm*

Why Using PPG?

Automatic detection

Causes slow variation in DC component

Introduces 3 types of modulation

5 of 20

Thesis Overview

*bpm=breaths per minute

What is Respiratory Rate (RR)?

number of breaths you take per minute
normal RR for an adult at rest is 12 to 20 bpm*

Why Using PPG?

Automatic detection

Causes slow variation in DC component

Introduces 3 types of modulation

Less erroneous than manual techniques

Can be used as wearable device

6 of 20

Thesis Overview

*bpm=breaths per minute

What is Respiratory Rate (RR)?

number of breaths you take per minute
normal RR for an adult at rest is 12 to 20 bpm*

Why Using PPG?

Automatic detection

Causes slow variation in DC component

Introduces 3 types of modulation

Less erroneous than manual techniques

Can also be acquired using mobile cameras

Can be used as wearable device

7 of 20

Conventional Methods

Extraction of Respiratory signals

RR Estimation

Fusion of RR estimates

Raw PPG Signals

RR

Filter based

Feature based

8 of 20

Conventional Methods

Extraction of Respiratory signals

RR Estimation

Fusion of RR estimates

Raw PPG Signals

RR

Freq. based

Time

based

9 of 20

Conventional Methods

Extraction of Respiratory signals

RR Estimation

Fusion of RR estimates

Raw PPG Signals

RR

Modulation

Temporal

10 of 20

Conventional Methods

Extraction of Respiratory signals

RR Estimation

Fusion of RR estimates

Raw PPG Signals

RR

Limitations

Computational Complexity
Less Accuracy
Not suitable for real time application
Results can vary due to sensor model

11 of 20

Proposed Algorithm

cnn_model

LSTM

ANN

CNN

Input Vector

Input (Clean PPG)

Output RR

12 of 20

Proposed Algorithm

cnn_model

LSTM

ANN

CNN

ANN

Input Vector

To LSTM

Input (Clean PPG)

Output RR

13 of 20

Proposed Algorithm

cnn_model

LSTM

ANN

Input Vector

Output

ReLu

Input (Clean PPG)

Output RR

Fully Connected Layer

14 of 20

Preprocessing

Raw PPG Signal

Filtered Signal

Normalized Signal

1^st Derivative

Signal with detected peaks & troughs

Peak to peak Interval

Removing Corrupted Segments

15 of 20

Data Sources

MIMIC II Database
31 patients
PPG signal sampled at 125Hz
RR recorded at 1s interval
Total 31158 segments found

Vortal Database
10 patients
PPG signal sampled at 500Hz
Resampled at 125 Hz
RR recorded at 1s interval
Total 6132 segments found

16 of 20

Results: Convergence Analysis & Prediction

Training Loss= 0.13

Validation Loss=0.16

No of Epochs=104

Training split=70%

Validation split=10%

Testing Split=20%

17 of 20

Results: Statistical Analysis

Box-Plot for RR

Regression Plot for RR

Bland-Altman Plot for RR

The limits of agreement is [-3.18,3.08]

r=0.95

18 of 20

Results : Comparison

Table: Error Analysis

Model	Percentage of prediction for error
Model	<1 bpm	<2bpm	<3bpm
FTS[1]	85.4%	91.4%	93.31%
ARS[2]	86.2%	92.25%	94.68%
Proposed Model	87.01%	92.59%	94.66%

[1]Karlen et al.2013 Multiparameter respiratory rate estimation.

[2] Thayer et al.2002 Estimating respiratory frequency.

19 of 20

Future Work

Generalize the algorithm for detection under exercise and other clinical condition
Implementation in smart devices

20 of 20

Thank you