Academia.eduAcademia.edu

Figure 4 – uploaded by Martin Glavin

See full PDFdownloadDownload figure
Fig. 4. Boxplots of SD-Q6 (a) and AD-Q6 (b) showing the probability of a "good" reconstruction.

Figure 4 Boxplots of SD-Q6 (a) and AD-Q6 (b) showing the probability of a "good" reconstruction.

Related Figures (13)

Fig. 1. Examples of 20 random training signals for various dictionary creations. (a) Standard DL approach where all types of signals are used in training to create Y;. (b) Training frames for a sub-dictionary Y%yg where no QRS complex is detected in a frame. (c) Training frames for a sub-dictionary Woo where q = 12 and N = 192. The red box indicates the defined region where the QRS complex is present, for the training signals. are created YQ = [Yo Yo2,%o3...Woq]. Each dictionary is created with training signals that contain QRS complexes across individual temporal windows within the frame, each containing N/g samples, shifted across the frame with no overlap between windows. For example, the set of training signals used to create the dictionary Vg; contain QRS complexes in the sample regions [1:N/q]. In general, this concept allows the signal reconstruction quality to be improved, as the chosen sub-dictionary is trained for sparse representations of signals where the majority of signal energy in the frame is concentrated on that location. Fig. 1 demonstrates an example of training signals used to create the dictionaries ¥; (a), Myo (b) and Yo (c) (with g = 12 and N = 192).
Fig. 1. Examples of 20 random training signals for various dictionary creations. (a) Standard DL approach where all types of signals are used in training to create Y;. (b) Training frames for a sub-dictionary Y%yg where no QRS complex is detected in a frame. (c) Training frames for a sub-dictionary Woo where q = 12 and N = 192. The red box indicates the defined region where the QRS complex is present, for the training signals. are created YQ = [Yo Yo2,%o3...Woq]. Each dictionary is created with training signals that contain QRS complexes across individual temporal windows within the frame, each containing N/g samples, shifted across the frame with no overlap between windows. For example, the set of training signals used to create the dictionary Vg; contain QRS complexes in the sample regions [1:N/q]. In general, this concept allows the signal reconstruction quality to be improved, as the chosen sub-dictionary is trained for sparse representations of signals where the majority of signal energy in the frame is concentrated on that location. Fig. 1 demonstrates an example of training signals used to create the dictionaries ¥; (a), Myo (b) and Yo (c) (with g = 12 and N = 192).
VARIABLES USED IN POWER CALCULATIONS TABLE I With PRD, the mean of the signal X is not removed from the original signal. This can result in a DC bias artificially lowering the distortion results. PRD* is therefore more commonly used, however, PRD is included for more direct comparison with CS implementations in the literature.
VARIABLES USED IN POWER CALCULATIONS TABLE I With PRD, the mean of the signal X is not removed from the original signal. This can result in a DC bias artificially lowering the distortion results. PRD* is therefore more commonly used, however, PRD is included for more direct comparison with CS implementations in the literature.
Fig. 2. Block diagrams for the acquisition, compression and transmission operations for CS (top) and SPIHT (bottom). A Spread Spectrum Random Modulation Pre-Integrator (SRMPI) design is used for CS acquisition. SPIHT includes a Nyquist rate ADC. CS compression includes redundancy removal, quantization and Huffman coding before BLE wireless transmission. SPIHT comprises a DWT and SPIHT coding before the binary stream is transmitted. The Cs and sFlmt architectures assumed in this paper for acquisition, compression and transmission are detailed in Fig. 2. The power consumption of each algorithm is determined by calculating the power required in (i) sampling the signal, (ii) performing digital signal compression on a Blackfin (BF537) DSP [14] and (iii) wireless transmission of the ECG data using the Texas Instruments CC2540 BLE wireless transceiver [15]. The variables used in the power analysis are defined in Table I, and their values given. Each component of the power calculation is now treated in more detail.
Fig. 2. Block diagrams for the acquisition, compression and transmission operations for CS (top) and SPIHT (bottom). A Spread Spectrum Random Modulation Pre-Integrator (SRMPI) design is used for CS acquisition. SPIHT includes a Nyquist rate ADC. CS compression includes redundancy removal, quantization and Huffman coding before BLE wireless transmission. SPIHT comprises a DWT and SPIHT coding before the binary stream is transmitted. The Cs and sFlmt architectures assumed in this paper for acquisition, compression and transmission are detailed in Fig. 2. The power consumption of each algorithm is determined by calculating the power required in (i) sampling the signal, (ii) performing digital signal compression on a Blackfin (BF537) DSP [14] and (iii) wireless transmission of the ECG data using the Texas Instruments CC2540 BLE wireless transceiver [15]. The variables used in the power analysis are defined in Table I, and their values given. Each component of the power calculation is now treated in more detail.
Fig. 3. PRD* (a) and PRD (b) comparison between the SD and AD reconstruction approaches, the MMB-IHT and MMB-CoSAMP, and BSBL approaches in the literature and SPIHT. Q6 and Q7 indicate quantization of the difference values at 6 and 7 bits respectively. The proposed AD-Q6, SD-Q7, SPIHT and the results presented in [11] for MMB-IHT, MMB-CoSAMP and BSBL are compared in terms of PRD in Fig. 3(b). It is clear AD-Q6 outperforms the MMB methods and BSBL for all CRs tested. The results in [11] only provide values for PRD over a range of CRs and therefore these results are shown here instead of PRD*,. However, this should not change our general conclusions in terms of comparison with these CS methods. While SPIHT still gives better overall performance, among the CS-based methods, the results obtained with the AD approach proposed here demonstrate a significant improvement in CS ECG implementations.
Fig. 3. PRD* (a) and PRD (b) comparison between the SD and AD reconstruction approaches, the MMB-IHT and MMB-CoSAMP, and BSBL approaches in the literature and SPIHT. Q6 and Q7 indicate quantization of the difference values at 6 and 7 bits respectively. The proposed AD-Q6, SD-Q7, SPIHT and the results presented in [11] for MMB-IHT, MMB-CoSAMP and BSBL are compared in terms of PRD in Fig. 3(b). It is clear AD-Q6 outperforms the MMB methods and BSBL for all CRs tested. The results in [11] only provide values for PRD over a range of CRs and therefore these results are shown here instead of PRD*,. However, this should not change our general conclusions in terms of comparison with these CS methods. While SPIHT still gives better overall performance, among the CS-based methods, the results obtained with the AD approach proposed here demonstrate a significant improvement in CS ECG implementations.
Fig. 5. Examples of two PVC events in record 119. The original signals are in blue and the reconstructed signals (with a CR = 17.6) are in red. (a) SPIHT reconstruction with a PRD* of 7.05%; (b) AD reconstruction with a PRD* of 6.66%. The black circle indicates the location of the annotated PVCs. Boxplots for SD-Q6 (a) and AD-Q6 (b) show. th probability of "good reconstruction" per 1024 sample: averaged over all records in Fig. 4. In each box the median i: the central line, the box edges signify the 25th and 75th percentiles and the whiskers show the most extreme dat points that are not considered outliers. It is clear from Fig. 4 that the AD improves the probability of "good" reconstruction t also reduces the variance in reconstruction performanc« from record to record.
Fig. 5. Examples of two PVC events in record 119. The original signals are in blue and the reconstructed signals (with a CR = 17.6) are in red. (a) SPIHT reconstruction with a PRD* of 7.05%; (b) AD reconstruction with a PRD* of 6.66%. The black circle indicates the location of the annotated PVCs. Boxplots for SD-Q6 (a) and AD-Q6 (b) show. th probability of "good reconstruction" per 1024 sample: averaged over all records in Fig. 4. In each box the median i: the central line, the box edges signify the 25th and 75th percentiles and the whiskers show the most extreme dat points that are not considered outliers. It is clear from Fig. 4 that the AD improves the probability of "good" reconstruction t also reduces the variance in reconstruction performanc« from record to record.
Fig. 6. Examples of three original and reconstructed signals from record 119 for different levels of SNR. The original signals are illustrated in blue and the reconstructed signals are in red. (a) shows an original clean signal and a reconstructed signal with a PRD* of 9.3% at a CR of 10.97, (b) shows a signal with a SNR of 18 dB reconstructed with a PRD* of 7.12% at a CR of 10.24 and (c) shows a signal with a SNR of 12 dB reconstructed with a PRD* of 6.81% at a CR of 9.54. dB SNR levels. It is clear that the signals present more variability and a more difficult reconstruction problem for the proposed algorithm.
Fig. 6. Examples of three original and reconstructed signals from record 119 for different levels of SNR. The original signals are illustrated in blue and the reconstructed signals are in red. (a) shows an original clean signal and a reconstructed signal with a PRD* of 9.3% at a CR of 10.97, (b) shows a signal with a SNR of 18 dB reconstructed with a PRD* of 7.12% at a CR of 10.24 and (c) shows a signal with a SNR of 12 dB reconstructed with a PRD* of 6.81% at a CR of 9.54. dB SNR levels. It is clear that the signals present more variability and a more difficult reconstruction problem for the proposed algorithm.
ACQUISITION POWER PROFILES FOR SPIHT AND CS
ACQUISITION POWER PROFILES FOR SPIHT AND CS
AVERAGE CLOCK CYCLES PER 1s FOR EACH OPERATION TABLE II
AVERAGE CLOCK CYCLES PER 1s FOR EACH OPERATION TABLE II
Fig. 7. Boxplots of the PRD* of AD-Q6 and SD-Q6 tested at different SNR levels for approximately equal CRs. The performance at (a) 12 dB (b) 18 dB (c) 24 dB and (d) the clean signals are illustrated. performance over SD-Q6 for increasing SNR levels is important. The AD-Q6 implementation outperforms the SD- Q6 at each tested CR. It is important to note that the AD algorithm maintains improved performance over the SD approach, and the implementation is robust to the presence of the noise related artefacts and electrode motion in reconstruction to demonstrate the full potential of the method for implementation in low-power BANs.
Fig. 7. Boxplots of the PRD* of AD-Q6 and SD-Q6 tested at different SNR levels for approximately equal CRs. The performance at (a) 12 dB (b) 18 dB (c) 24 dB and (d) the clean signals are illustrated. performance over SD-Q6 for increasing SNR levels is important. The AD-Q6 implementation outperforms the SD- Q6 at each tested CR. It is important to note that the AD algorithm maintains improved performance over the SD approach, and the implementation is robust to the presence of the noise related artefacts and electrode motion in reconstruction to demonstrate the full potential of the method for implementation in low-power BANs.
SS Se Oe ee eT ee Fig. 8. Power profiles of individual components for each compression algorithm. The power consumption of each process is shown for SPIHT and CS. Symbols are defined in Section IV. used). At a CR of 2, CS consumes 4.73 uW and Pacs decreases as compression increases. This is due to a reduction in the number of measurements (M) required which reduces Papc.
SS Se Oe ee eT ee Fig. 8. Power profiles of individual components for each compression algorithm. The power consumption of each process is shown for SPIHT and CS. Symbols are defined in Section IV. used). At a CR of 2, CS consumes 4.73 uW and Pacs decreases as compression increases. This is due to a reduction in the number of measurements (M) required which reduces Papc.
Fig. 9. Total system power profiles of the AD-Q6 implementation, SPIHT, MMB-IHT, MMB-CoSAMP and BSBL for increasing PRD value. F exis or all PRD values considered, AD outperforms the other ting CS techniques. for PRD values > 3%, AD-Q6 is the best performing algorithm. However, SPIHT performs best in terms of power consumption for PRDs less than 3%. The dec ine in reconstruction quality for increasing CRs is sharper for the MMB and BSBL techniques and SPIHT outperforms MM B-CoSAMP and BSBL at all PRDs shown despite the advantages in lower power sampling and compression operations. For PRD values > 4.5% MMB-IHT outperforms SPI HT for the PRD range shown. It is clear the performance of t he AD is superior to the CS techniques tested and this appears to increase as the level of PRD increases.
Fig. 9. Total system power profiles of the AD-Q6 implementation, SPIHT, MMB-IHT, MMB-CoSAMP and BSBL for increasing PRD value. F exis or all PRD values considered, AD outperforms the other ting CS techniques. for PRD values > 3%, AD-Q6 is the best performing algorithm. However, SPIHT performs best in terms of power consumption for PRDs less than 3%. The dec ine in reconstruction quality for increasing CRs is sharper for the MMB and BSBL techniques and SPIHT outperforms MM B-CoSAMP and BSBL at all PRDs shown despite the advantages in lower power sampling and compression operations. For PRD values > 4.5% MMB-IHT outperforms SPI HT for the PRD range shown. It is clear the performance of t he AD is superior to the CS techniques tested and this appears to increase as the level of PRD increases.
SP Se ee Fig. 10. Comparison of two use cases presented with MMB-IHT and SPIHT over various time durations at an average PRD of 5.5%. Here we present two use case scenarios for training and analyze the effects on long term monitoring by estimating the average power consumption over various time periods. Case | includes the ideal scenario where training is performed in a clinical environment prior to implementation, adding no power costs to the BAN. Case 2 is the scenario where training is performed on the device itself. The architecture considered in this paper requires a training phase involving the collection of 15 minutes of ECG data. Therefore, Case 2 assumes Nyquist rate sampling for 15 minutes prior to the proposed AD implementation being employed. For comparison purposes, the two use cases will be compared with the MMB- THT algorithm and SPIHT. It should be noted the MMB approaches utilize Huffman dictionaries in _ their implementation which require offline creation. However, this is not included in our analysis as training periods are not specified in [6, 11] and we assume straightforward implementation of the MMB-IHT technique. Similarly, there is no requirement for a training period for SPIHT so the average power consumption stays constant in that case. At a CR of 14.8 the AD-Q6 method achieves an average PRD* of 9% (deemed acceptable for clinical inspection in [33]) and PRD of 5.5%. As only average PRD (rather than PRD*) values for were reported MMB-IHT [11] (around the target PRD*), a PRD of 5.5% will be considered the operating point for this use case analysis. At the target PRD of 5.5%, AD-Q6 consumes 30.9 uW, SPIHT consumes 50.65 uW and MMB- IHT consumes 49.4 wW.
SP Se ee Fig. 10. Comparison of two use cases presented with MMB-IHT and SPIHT over various time durations at an average PRD of 5.5%. Here we present two use case scenarios for training and analyze the effects on long term monitoring by estimating the average power consumption over various time periods. Case | includes the ideal scenario where training is performed in a clinical environment prior to implementation, adding no power costs to the BAN. Case 2 is the scenario where training is performed on the device itself. The architecture considered in this paper requires a training phase involving the collection of 15 minutes of ECG data. Therefore, Case 2 assumes Nyquist rate sampling for 15 minutes prior to the proposed AD implementation being employed. For comparison purposes, the two use cases will be compared with the MMB- THT algorithm and SPIHT. It should be noted the MMB approaches utilize Huffman dictionaries in _ their implementation which require offline creation. However, this is not included in our analysis as training periods are not specified in [6, 11] and we assume straightforward implementation of the MMB-IHT technique. Similarly, there is no requirement for a training period for SPIHT so the average power consumption stays constant in that case. At a CR of 14.8 the AD-Q6 method achieves an average PRD* of 9% (deemed acceptable for clinical inspection in [33]) and PRD of 5.5%. As only average PRD (rather than PRD*) values for were reported MMB-IHT [11] (around the target PRD*), a PRD of 5.5% will be considered the operating point for this use case analysis. At the target PRD of 5.5%, AD-Q6 consumes 30.9 uW, SPIHT consumes 50.65 uW and MMB- IHT consumes 49.4 wW.
Related topics:
Computer ScienceAlgorithmsArtificial IntelligenceMachine LearningMedicineWireless TechnologySignal CompressionLossy CompressionSignal ReconstructionCompression Ratio
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved