A Signal Processing Approach to the Detection of Pulmonary Edema - - PowerPoint PPT Presentation

A Signal Processing Approach to the Detection of Pulmonary Edema

EEE24B

Nicole Lim Sze Ting Dr Ser Wee

Pulmonary Edema

the accumulation of excess fluid in the lungs

Limitations of current methods

Accuracy dependent on clinician's experience Costly, require bulky machines Time-consuming Exposure to radiation

Proposed method: ML algorithm

Faster More consistent and reliable Portable; Suitable for ambulatory use

ACC 95.7% with 5 features via kNN (k = 3)

Previous studies

Yang, F., Ser, W., Yu, J., Foo, D., Yeo, D., Chia, P., & Wong,

• J. (n.d.). Lung Water

Detection using Acoustic Techniques (Rep.).

AIM:

Improve existing algorithm

Methodology

1

Data Collection & Pre-Processing

40s

Audio recordings of lung sounds are labelled by a doctor.

1

Data Collection & Pre-Processing

Recordings are divided into 3 second samples.

40s

40s 40s 3s

1

Data Collection & Pre-Processing

40s 40s 3s

1024

x x x x x x x x x x x x

1024

x x x x x x x x x x x x

1024

x x x x x x x x x x x x

1024

x x x x x x x x x x x x

Samples are divided into windows

• f 1024 points, with 50% overlap.

1

Data Collection & Pre-Processing

100

y x

t1 (t1, y1) (t1, ✕ 100) y1 ymax

Windows are normalised such that values range from 0 to 100.

3s

fn

• No. of windows

= ∑

Average feature value of windows gives final feature value of sample.

1024

x x x x x x x x x x x x

fn

1024

x x x x x x x x x x x x

fn

1024

x x x x x x x x x x x x

fn

1024

x x x x x x x x x x x x

fn

Feature Extraction

2

3

Feature Selection

(meanA - meanB) 2 varianceA + varianceB Fisher’s ratio =

Fisher's Ratio

Higher FR Lower FR

Feature Extraction

2 Features used in the detection of wheezing Features previously used in the detection of PE

Aydore, S., Sen, I., Kahya, Y., & Mihcak, M. (n.d.). Classification of Respiratory Signals by Linear Analysis (Rep.).

Feature Extraction

2 Previous algorithms for the detection of wheezing

Feature Extraction

2 Features used in the detection of wheezing

Kurtosis

Degree of peakedness of distribution

Renyi Entropy

Randomness of system

Mean Crossing Irregularity & Frequency

Mean-crossing behaviour

Feature Extraction

2

Fisher's Ratio of features used in the detection of wheezing

Kurtosis Renyi Entropy MCI MCF

0.0010 0.0053 0.0086 0.0349

Feature Extraction

2 Features previously used in the detection of PE

13 Mel-Frequency Cepstral Coefficients (MFCCs)

Mimic doctor’s logarithmic perception of lung sounds during auscultation

Feature Extraction

2 Features previously used in the detection of PE

13 Mel-Frequency Cepstral Coefficients (MFCCs)

Mel scale:

Approximated frequency resolution

• f the human auditory

system

Mels kHz

Feature Extraction

2 Features previously used in the detection of PE

Feature Extraction

2 Features previously used in the detection of PE

Ratio & Difference between MFCCs

• Hypothesised to have higher

discriminating power

• Derived from top 6 MFCC

Feature Extraction

2 Features previously used in the detection of PE

Using Fisher's Ratio in Feature Selection 3

Feature Selection

• Features with higher FR values

are added first

○ Reduce number of features ○ Higher classification accuracy ○ Shorten training time

3

Feature Selection

Final feature ranking by FR

Signal Classifiers

k-Nearest Neighbours (kNN)

Most common label amongst k nearest neighbours

Support Vector Machines (SVM)

Decision boundary/hyperplane determined by support vectors

Signal Classification

4

Evaluation metrics

Model Evaluation

5

TPR

True positive rate

TNR

True negative rate

ACC

Detection accuracy

10-fold cross validation

Evaluation metrics

Model Evaluation

5

ACC ACC

Healthy Unhealthy

Algorithm

Healthy Unhealthy

Doctor

Evaluation metrics

Model Evaluation

5

Missed diagnosis

TPR

Healthy Unhealthy

Algorithm

Healthy Unhealthy

Doctor

Evaluation metrics

Model Evaluation

5

False alarms TNR

Healthy Unhealthy

Algorithm

Healthy Unhealthy

Doctor

Evaluation metrics

Model Evaluation

5

TPR

True positive rate

TNR

True negative rate

ACC

Detection accuracy

10-fold cross validation

10-fold cross validation

Model Evaluation

5

training training training training training training training training training training testing

10-fold cross validation

Model Evaluation

5

training training training training training training training training training training testing testing testing testing testing testing testing testing testing testing

Removing features that cause decrease in ACC

Performance Improvement

6

Removing features that cause decrease in ACC

Performance Improvement

6

• 1. Rank features by decrease in

ACC caused

• 2. Remove features from the

model in order of descending decrease caused

Results & Discussion

Best Model: kNN

ACC

86

TPR

89

TNR

82

kNN (k = 1)

• n 24 features

Best Model: kNN

Effect of removing 6 features

TPR

89 88

ACC

86 85

TNR

82 80 24 18 24 18 24 18

Best Model: SVM

SVM (C = 1)

• n 26 features

ACC

86

TPR

88

TNR

82

Best Model: SVM

Effect of removing 9 features

TPR

88 88

ACC

86 86

TNR

82 83 26 17 26 17 26 17

Comparison to results of previous works

kNN

ACC

85 85 13 18

TPR

85 88 13 18

TNR

85 80 13 18 YF ME

Comparison to results of previous works

SVM

TPR

88 88

ACC

87 86

TNR

84 83 12 17 12 17 12 17 YF ME

Comparison to results of previous works

• New features with higher FR

values via ratio/difference of MFCC

• Did not improve ACC

Evaluation of my approach

Strategies hypothesised to improve ACC

Strategy Evaluation

Features used for the detection of other breathing anomalies Derivation of features using ratio and difference Removal of features that decrease ACC to improve performance

Strategies hypothesised to improve ACC

Features used for the detection of

• ther breathing anomalies
• Not useful in distinguishing

healthy and unhealthy signals ○ PE: crackle sounds

Strategy Evaluation

Strategies hypothesised to improve ACC

Derivation of features using ratio and difference

• Can derive features with higher

FR ⇒ ↑ discriminating power

• Time-consuming

○ Comparison of boxplots

Strategy Evaluation

Strategies hypothesised to improve ACC

Renyi entropy

Randomness of system

Mean Crossing Irregularity (MCI) & Mean Crossing Frequency (MCF)

Mean-crossing behaviour

Removal of features that decrease ACC to improve performance

• ↓ training time
• ↓ algorithm complexity
• Can potentially improve ACC

○ SVM algorithm

Strategy Evaluation

Limitations

Uncertainty in performance evaluation due to validation method Quality of data & reliability of doctor

Future Work

Automate derivation of ratio/ difference-based features Vary number of MFCCs

Conclusion

Yang Feng's algorithm performed better Proposed algorithm

• SVM (C = 1)
• 17 features
• ACC 85.8, TPR 88, TNR 83

Key points

SVM > kNN

TPR

88

ACC

85

TNR

80 24 18 24 18 24 18

Key points

SVM kNN

88 86 83 17 17 17

Wheeze detection features for pulmonary edema detection

Key points

Difference/ratio-based feature derivation for features with higher Fisher's ratio values Removing features that decrease accuracy to improve performance

Dr Ser Wee

Research mentor

Acknowledgements

Dr Shi Wen

Advice on MATLAB and data handling

Mr Low Kay Siang

Teacher-mentor

Q&A