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

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

Faster Proposed method: More consistent and reliable ML algorithm Portable; Suitable for ambulatory use

Yang, F., Ser, W., Yu, J., Foo, D., Yeo, D., Chia, P., & Wong, J. (n.d.). Lung Water Detection using Acoustic Previous Techniques (Rep.). studies ACC 95.7% with 5 features via kNN (k = 3)

AIM: Improve existing algorithm

Methodology

40s 1 Data Collection & Audio recordings of lung sounds Pre-Processing are labelled by a doctor.

1 40s 40s 40s 3s Data Collection & Recordings are divided into 3 Pre-Processing second samples.

x x x x x x x x x x x x x x x x 1 x x x x x x x x x x x x 1024 40s x x x x x x x x x x x 1024 x x x x x x x x 40s 1024 x 1024 3s Data Collection & Samples are divided into windows of 1024 points, with 50% overlap. Pre-Processing

y 100 (t 1 , y 1 ) 1 y 1 (t 1 , ✕ 100) y max Data Collection & x t 1 Pre-Processing Windows are normalised such that values range from 0 to 100.

x x x x x f n x x x x x x x x x x = ∑ x f n x x x x x x x x f n x x x x 1024 f n x 2 x x x x x x x x x x 1024 f n x x x x x x x x 1024 x 1024 3s No. of windows Feature Average feature value of windows Extraction gives final feature value of sample.

Fisher's Ratio (mean A - mean B ) 2 Fisher’s ratio = variance A + variance B 3 Feature Selection Higher FR Lower FR

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

Previous algorithms for the detection of wheezing Aydore, S., Sen, I., Kahya, Y., & 2 Mihcak, M. (n.d.). Classification of Respiratory Signals by Linear Feature Analysis (Rep.). Extraction

Features used in the detection of wheezing Kurtosis 2 Degree of peakedness of distribution Feature Renyi Entropy Extraction Randomness of system Mean Crossing Irregularity & Frequency Mean-crossing behaviour

Fisher's Ratio of features used in the detection of wheezing Kurtosis 0.0010 2 Renyi 0.0053 Feature Entropy Extraction MCI 0.0086 MCF 0.0349

Features previously used in the detection of PE 13 Mel-Frequency Cepstral 2 Coefficients (MFCCs) Feature Mimic doctor’s logarithmic Extraction perception of lung sounds during auscultation

Features previously used in the detection of PE 13 Mel-Frequency Cepstral 2 Coefficients (MFCCs) Feature Mels Mel scale: Extraction Approximated frequency resolution of the human auditory system kHz

Features previously used in the detection of PE 2 Feature Extraction

Features previously used in the detection of PE Ratio & Difference between MFCCs 2 ● Hypothesised to have higher Feature discriminating power Extraction ● Derived from top 6 MFCC

Features previously used in the detection of PE 2 Feature Extraction

Using Fisher's Ratio in Feature Selection ● Features with higher FR values 3 are added first Feature ○ Reduce number of features Selection ○ Higher classification accuracy ○ Shorten training time

Final feature ranking by FR 3 Feature Selection

Signal Classifiers k-Nearest Neighbours (kNN) 4 Most common label amongst k nearest neighbours Signal Classification Support Vector Machines (SVM) Decision boundary/hyperplane determined by support vectors

Evaluation metrics ACC 5 Detection accuracy 10-fold TPR Model cross True positive rate Evaluation validation TNR True negative rate

Evaluation metrics Algorithm Healthy Unhealthy 5 Unhealthy ACC Model Doctor Evaluation Healthy ACC

Evaluation metrics Algorithm Healthy Unhealthy 5 Unhealthy Missed TPR diagnosis Model Doctor Evaluation Healthy

Evaluation metrics Algorithm Healthy Unhealthy 5 Unhealthy Model Doctor Evaluation False Healthy TNR alarms

Evaluation metrics ACC 5 Detection accuracy 10-fold TPR Model cross True positive rate Evaluation validation TNR True negative rate

10-fold cross validation training testing training training 5 training training training Model Evaluation training training training training

10-fold cross validation training testing training testing training testing 5 training testing training testing training testing Model Evaluation training testing training testing training testing testing training

Removing features that cause decrease in ACC 6 Performance Improvement

Removing features that cause decrease in ACC 6 1. Rank features by decrease in ACC caused Performance 2. Remove features from the Improvement model in order of descending decrease caused

Results & Discussion

89 86 82 kNN (k = 1) Best Model: on 24 features kNN ACC TPR TNR

Effect of removing 6 features 89 88 86 85 82 80 Best Model: kNN 24 18 24 18 24 18 ACC TPR TNR

88 86 82 SVM (C = 1) Best Model: on 26 features SVM ACC TPR TNR

Effect of removing 9 features 88 88 86 86 83 82 Best Model: SVM 26 17 26 17 26 17 ACC TPR TNR

kNN YF ME 88 85 85 85 85 80 Comparison to results of previous works 13 18 13 18 13 18 ACC TPR TNR

SVM YF ME 88 88 87 86 84 83 Comparison to results of previous works 12 17 12 17 12 17 ACC TPR TNR

Evaluation of my approach ● New features with higher FR values via ratio/difference of Comparison to MFCC results of previous ● Did not improve ACC works

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

Strategies hypothesised to improve ACC Features used for the detection of other breathing anomalies Strategy ● Not useful in distinguishing Evaluation healthy and unhealthy signals ○ PE: crackle sounds

Strategies hypothesised to improve ACC Derivation of features using ratio and difference Strategy ● Can derive features with higher Evaluation FR ⇒ ↑ discriminating power ● Time-consuming ○ Comparison of boxplots

Strategies hypothesised to improve ACC Renyi entropy Removal of features that decrease Randomness of system ACC to improve performance Strategy Mean Crossing ● ↓ training time Evaluation ● ↓ algorithm complexity Irregularity (MCI) & ● Can potentially improve ACC Mean Crossing ○ SVM algorithm Frequency (MCF) Mean-crossing behaviour

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

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

Conclusion

Proposed algorithm ● SVM (C = 1) ● 17 features ● ACC 85.8, TPR 88, TNR 83 Key points Yang Feng's algorithm performed better

SVM > kNN SVM kNN 88 88 86 85 83 80 Key points 24 17 18 24 17 18 24 17 18 ACC TPR TNR

Wheeze detection features for pulmonary edema detection Difference/ratio-based feature derivation for features with higher Fisher's ratio values Key points Removing features that decrease accuracy to improve performance

Dr Ser Wee Research mentor Dr Shi Wen Advice on MATLAB and data handling Acknowledgements Mr Low Kay Siang Teacher-mentor

Q&A