Abstract
Design of an embedded electronic system, the "BodyGateWay" (Fig. 1), equipped with miniaturised silicon sensors and circuits for the detection of biologic signals (tri-axis accelerometer and filters/amplifiers for electrocardiography), a low-power microcontroller for real-time on-board processing, wireless Bluetooth transmission module.
The processing of the electrocardiogram includes: noise and artifacts removal, heart rate estimation, respiratory rate estimation, heart beat classification, heart rhythm evaluation.
The processing of the accelerometer data includes digital filters for noise removal, physical activity level estimation by means of a step counter. Signals are processed not only independently but also jointly: e.g. correspondence between heart rate and physical activity is verified.
Fig.1 Several BodyGateWay prototypes
1. INTRODUCTION
Miniaturised sensors allow continuous comfortable and precise detection of biologic signals. Significant parameters are extracted by several algorithms based on this signals. These parameters can then be used for diagnosis and be checked by medical doctors.
The embedded system is low-cost and increases the efficiency of any care protocol. As an example, a wider population can be monitored with the same amount of expenses. This is particularly important as the population is becoming older.
There are several factors to be considered. First, the device has to be small and light, so that it can be worn with comfort. That is why the device must be miniaturised. Also, it must be robust to stress while it is worn.
Second, it must allow the detection of biologic signals for a long period of time. That is why the device must be low-power.
Third, it must include a processing unit to clean up detected signals, to compress them (to consume less power when transmitting) and to process them (to extract significant parameters).
The device must include a module to transmit detected data for remote storage in order to enable further post-processing. To reach all this goals, the device can use several silicon sensors: electronic sensors for voltage, impedance and temperature; optical sensors (LEDs and photodiodes); mechanical sensors (MEMS accelerometers and gyroscopes, micro electro-mechanical systems).
The current prototype does include voltage detection for the electrocardiogram, the impedance detection (current injection and again voltage detection), 3-axis acceleration detection. The prototype under construction will include two LEDs and one photodiode for the photoplethysmography, two temperature sensors (one for the skin and one for the ambient), one microphone for internal sounds (cardiac and respiration sounds) and external sounds (voice).
The low-power microcontroller manages everything from data acquisition to data processing, local/remote storage, transmission/reception via Bluetooth.
Reception of data is obtained after pairing with external sensor such as scales and blood pressure cuffs. Transmission of data is obtained by pairing with devices such as smartphones and personal computers. A rechargeable battery powers the whole system.
The device is small, light and somewhat flexible. It can be worn as a patch. The current prototype has been designed to be attached on the sternum, in the middle of the chest. Biologic signals collected from on-board sensors and external sensors connected via radio are processed and retransmitted to local devices (smartphone) which in turn retransmit everything to remote devices (servers) dedicated to data storage and retrieval (Fig. 2).
Fig.2 Complete system featuring the electronic patch along different off-body sensors. The device is designed to detect biological signals, process and transmit them via Bluetooth to a nearby mobile phone, who in turns sends them to a remote healthcare server, where the data can be easily reviewed using a web browser.
2. MULTIPLE SENSORS
The device does include two different kinds of sensors. One kind is dedicated to voltage detection for the electrocardiogram and bio-impedance. The other kind is dedicated to motion detection by means of a 3-axis accelerometer. Signals are first processed independently to clean them up and extract significant parameters. Then signals and parameters are processed jointly.
2.1 Electrocardiogram detection
The electrocardiogram is detected by amplifying the voltage difference between two electrodes placed on the chest 8-10 cm apart from each other. The electronic circuit cancels the common mode voltage, bandpass filters the signal (band of interest goes from 0.05-2 Hz to 100 Hz). Sampling is performed at 500 Hz, but the sampling frequency can be programmed and set from 100 Hz to 4 kHz. Quantization is done with 12 bits of precision.
2.2 Motion detection
The motion is detected by means of a 3-axis accelerometer. The signals are sampled at 25 Hz with 8 bits of resolution. A proprietary algorithm extract the number of steps, estimates the instantaneous body velocity and the distance travelled. The algorithm is designed to work with the accelerometer placed on the chest (as it is done in our prototype), on the arm and on the waist.
2.3 Bio-impedance detection
The detection of the impedance is done by using the same electrodes used by the electrocardiogram. A continuous or alternated current is injected into the body, the voltage is measure, the ratio being the impedance. Several seconds are required to switch from the electrocardiogram detection to bio-impedance detection in order to allow transients in analog filters to complete.
2.4 Altri sensori
The device will include other sensors. One microphone to detect cardiac sound (phonocardiogram) and respiratory sounds. The signal is sampled a 4 kHz with 16 bits resolution. One CMOS photodiode to detect the light reflect by tissues under the skin. The two LEDs emit light in the red and infra-red band. The signal is used for the photoplethysmogram (oxygen saturation estimation).
3. INDEPEDENT PROCESSING
The software which is run on the microcontroller is designed to filter signals coming from each sensor. The first step in the processing chain does include an independent filter chain for each signal.
The goal of the independent processing is to remove the noise and detect artifacts. Out of band noise is removed by bandpass filtering. Cleaned signals are used to estimate significant parameters (heart rate, respiratory rate, etc). Artifacts (false signals generated by motion) and anomalous conditions (electrode disconnection) are detected, obviously in this case the parameters are not available.
Parameter extraction is prone to false negatives (pattern is present but is not detected) and false positive (pattern detected even if not present). A high sensitivity would result in the correct detection of all patterns present in the signal but would also result in an higher false positive rate.
In the case of biomedical applications, it is preferred to have a high precision, to be sure that when the pattern is declared detected it is really present into the signal. Unfortunately this results in a higher false negative rate.
3.1 Electrocardiogram signal processing
Two different processing chains were implemented and evaluated. The first one is the Hamilton-Tompkins [1][2] processing chain, used by a well-known publicly available algorithm for heart rate estimation (OSEA). The chain consists in the following blocks: a low-pass filter with a 16 Hz cut-off frequency that eliminates high-frequency noise (muscular noise and 50-60 Hz power-line interference), a highpass filter with a 8 Hz cut-off frequency which removes baseline wander, a block that computes the first-order derivative (as peaks corresponding to heart beats generally present considerable slopes), a 80 ms moving average filter which highlights peaks, and an adaptive threshold detection logic.
The second chain is based on the Pan-Tompkins algorithm, and it has been chosen for the BGW prototype. While the filtering chain is essentially the same, the slope is squared and not taken in absolute value, and the moving average is performed using a 150 ms window instead of a 80 ms one. This results in a more effective peak detection, even in the case of low-amplitude peaks, in spite of a greater computational effort.
Fig.3 Results of the electrocardiogram processing performer by the Pan-Tompkins chain and the simple adaptive-threshold logic for the detection of peaks related to heartbeats. Red: false positives (events detected but not present). Green: false negatives (undetected events which instead took place). Black: Correct classification but on the wrong part of the waveform (e.g. P wave instead of QRS complex).
The first experimental results show two critical situations: electrodes disconnection and noise overlapping the spectral bandwidth of the signal of interest. They both lead to an erroneous detection of R wave peaks, with a consequent wrong evaluation of the heart rate (Fig. 3). The algorithm was improved in three ways. First, an evaluation of the signal variance allows the detection of critical events (e.g. electrodes disconnection, high-frequency noise), because the variance of the signal during these events is significantly higher. The second improvement is based on the computation of the similarity (matching) with a reference waveform (template). The similarity is evaluated by detecting the local minimums of the sum of the absolute differences (SAD) between the current signal portion and the template [3]. Finally, a template database is used to make the algorithm independent of the acquired waveform - this is necessary because the prototype must function regardless of the electrodes placement.
Fig.4 Processing chain and detection logic for heartbeats (R wave peaks)
Templates are updated according to recognized patterns. This grants an adaptation to the specific ECG waveform of the patient. The detection logic is showed in Fig. 4, the template database is instead presented in Fig. 5 with the results of the processing (to be compared with Fig. 3 in order to understand the performance improvement).
Fig.5 (left) database of waveform templates used for the R-wave peak classification; (right) results from processing: in the upper plot, the classification performed by the standard reference algorithm (OSEA), which results in 10 false positives, 24 false negatives and the proper classification of 25/49 beats; in the lower plot the classification performed by the modified version of the algorithm (STM), which results in 0 false positives, 10 false negatives and the proper classification of 39/49 beats.
3.2 Accelerometer signal processing
The accelerometer signal is filtered and processed in order to extract a parameter corresponding to the total activity, counting steps and therefore measuring the physical activity level. The detection is performed by identifying steep transitions on the accelerometer axes (Fig. 6, on the left). During these events, the ECG signal is affected by noise (probably due to pectoral muscles balancing during the walk). A more sophisticated elaboration would permit the estimation of posture (prone, supine, standing), acts (walk, run) and anomalous events (falls).
Fig.6 Synchronised detection of electrocardiogram (upper waveform) and acceleration along three axis (x, y, z); (left) the patient is walking (noisy ECG); (center) patient standing still (heart rhythm present also on the xyz signals); (right) patient standind still but breathing heavily (breathing rhythm also on ECG and XYZ).
Heart and respiratory rates can be estimated from the accelerometer signal as well (Fig. 6, center and right). As a matter of fact, the accelerometer is capable of detecting weak signals with a frequency lower than 150 Hz, providing a good mechanomyography and a satisfying phonocardiogram. Unfortunately this detection is possible only when the patient is not moving. The respiratory rate can also be extracted from the ECG signal, by processing the R wave peaks envelope (Fig. 6, on the right).
A more sophisticated processing might allow the estimation of blood pressure from the acoustic/accelerometer signal [4].
3.3 Bio-impedance signal processing
The impedance signal acquired using a DC current allows the estimation of the electrical contact quality. The signal acquired using high-frequency currents (30-50 KHz) instead permits to extract the respiratory rate (however the electrodes optimal placement is different from the ECG one).
4. JOINT PROCESSING
After an independent processing, the prototype performs a joint processing.
4.1 Fusion of same parameters from different signals
The same parameter can be estimated from signals acquired using different sensors. A more reliable estimate can be obtained by combining signals from different channels. For example, the heart rate is primarily estimated using the electrocardiogram. The same signal can be confirmed by the results of the processing of the photoplethysmographic or the acoustic signal. When the patient is not moving, the accelerometer signal provides an useful estimate as well. Similarly the respiratory rate is detected mainly from the photoplethysmographic signal and from the impedance signal. The same signal can be confirmed by the microphone and, during stillness, by the accelerometer.
4.2 Comparison of different parameters from different signals
Since several parameters are available, it is possible to combine them to extract a more reliable information, or even a new one. For example, the photoplethysmographic signal is affected by noise artefacts that can be estimated with the accelerometer. The system performs a real-time cancellation of these artefacts, providing a more reliable information. In the same manner, the heart rate extracted from the electrocardiogram is compared with the estimate of the physical activity performed using the accelerometer, distinguishing between anomalous (stillness combined with tachycardia) and normal situations (running with a high heart rate).
5. PROTOCOL OF USE
The electronic patch is applied by paramedic personnel. In the start-up phase the correct detection of parameters is checked in different situations (e.g. stillness, walking or running). In this phase the software learns the typical waveforms and the typical values of the parameters of the specific situations. Therefore the device adapts to the specific patient. Afterwards the software performs a continuous monitoring, recording the parameters resulting from the processing. If anomalous events are recorded, the device records both the parameters and the raw signals, making them available for a subsequent examination. The software periodically acquires signals from other devices using the wireless connection (e.g. sent by an electronic scale) and sends the processing results to a central server.
6. BIOMEDICAL APPLICATIONS
The device has been conceived and designed to be used in a clinical context. However it is possibile to develop applications also in other fields: rehabilitation, training, healthcare.
7. BIBLIOGRAFIA
[1] B.-U. Kohler, C. Hennig and R. Orglmeister, "The principles of software QRS detection", IEEE Eng. Med. Biol. Mag., vol. 21, pp. 42-57, Jan./Feb. 2002.
[2] P. S. Hamilton and W. Tompkins, "Quantitative Investigation of QRS detection rules using the MIT/BIH arhythmia database", IEEE Trans Biomed. Eng., vol. BME-33, pp. 1157-1165, Dec. 1986.
[3] S.E. Dobbs, N. M. Schmitt, and H. S. Ozemek, "QRS detection by template matching using real-time correlation on a microcomputer", J. Clinic. Eng., vol 9, pp. 197-212, 1984.
[4] C. Tranulis, L. G. Durand, L. Senhadji and P. Pibarot, "Estimation of the Pulmonary Arterial Pressure by a Neural Network", Med. Biol. Eng. Comput. 03/2002, 40(2): 205-12.
--Andrea Vitali, Marco Pessione, STMicroelectronics
Design of an embedded electronic system, the "BodyGateWay" (Fig. 1), equipped with miniaturised silicon sensors and circuits for the detection of biologic signals (tri-axis accelerometer and filters/amplifiers for electrocardiography), a low-power microcontroller for real-time on-board processing, wireless Bluetooth transmission module.
The processing of the electrocardiogram includes: noise and artifacts removal, heart rate estimation, respiratory rate estimation, heart beat classification, heart rhythm evaluation.
The processing of the accelerometer data includes digital filters for noise removal, physical activity level estimation by means of a step counter. Signals are processed not only independently but also jointly: e.g. correspondence between heart rate and physical activity is verified.
Fig.1 Several BodyGateWay prototypes
1. INTRODUCTION
Miniaturised sensors allow continuous comfortable and precise detection of biologic signals. Significant parameters are extracted by several algorithms based on this signals. These parameters can then be used for diagnosis and be checked by medical doctors.
The embedded system is low-cost and increases the efficiency of any care protocol. As an example, a wider population can be monitored with the same amount of expenses. This is particularly important as the population is becoming older.
There are several factors to be considered. First, the device has to be small and light, so that it can be worn with comfort. That is why the device must be miniaturised. Also, it must be robust to stress while it is worn.
Second, it must allow the detection of biologic signals for a long period of time. That is why the device must be low-power.
Third, it must include a processing unit to clean up detected signals, to compress them (to consume less power when transmitting) and to process them (to extract significant parameters).
The device must include a module to transmit detected data for remote storage in order to enable further post-processing. To reach all this goals, the device can use several silicon sensors: electronic sensors for voltage, impedance and temperature; optical sensors (LEDs and photodiodes); mechanical sensors (MEMS accelerometers and gyroscopes, micro electro-mechanical systems).
The current prototype does include voltage detection for the electrocardiogram, the impedance detection (current injection and again voltage detection), 3-axis acceleration detection. The prototype under construction will include two LEDs and one photodiode for the photoplethysmography, two temperature sensors (one for the skin and one for the ambient), one microphone for internal sounds (cardiac and respiration sounds) and external sounds (voice).
The low-power microcontroller manages everything from data acquisition to data processing, local/remote storage, transmission/reception via Bluetooth.
Reception of data is obtained after pairing with external sensor such as scales and blood pressure cuffs. Transmission of data is obtained by pairing with devices such as smartphones and personal computers. A rechargeable battery powers the whole system.
The device is small, light and somewhat flexible. It can be worn as a patch. The current prototype has been designed to be attached on the sternum, in the middle of the chest. Biologic signals collected from on-board sensors and external sensors connected via radio are processed and retransmitted to local devices (smartphone) which in turn retransmit everything to remote devices (servers) dedicated to data storage and retrieval (Fig. 2).
Fig.2 Complete system featuring the electronic patch along different off-body sensors. The device is designed to detect biological signals, process and transmit them via Bluetooth to a nearby mobile phone, who in turns sends them to a remote healthcare server, where the data can be easily reviewed using a web browser.
2. MULTIPLE SENSORS
The device does include two different kinds of sensors. One kind is dedicated to voltage detection for the electrocardiogram and bio-impedance. The other kind is dedicated to motion detection by means of a 3-axis accelerometer. Signals are first processed independently to clean them up and extract significant parameters. Then signals and parameters are processed jointly.
2.1 Electrocardiogram detection
The electrocardiogram is detected by amplifying the voltage difference between two electrodes placed on the chest 8-10 cm apart from each other. The electronic circuit cancels the common mode voltage, bandpass filters the signal (band of interest goes from 0.05-2 Hz to 100 Hz). Sampling is performed at 500 Hz, but the sampling frequency can be programmed and set from 100 Hz to 4 kHz. Quantization is done with 12 bits of precision.
2.2 Motion detection
The motion is detected by means of a 3-axis accelerometer. The signals are sampled at 25 Hz with 8 bits of resolution. A proprietary algorithm extract the number of steps, estimates the instantaneous body velocity and the distance travelled. The algorithm is designed to work with the accelerometer placed on the chest (as it is done in our prototype), on the arm and on the waist.
2.3 Bio-impedance detection
The detection of the impedance is done by using the same electrodes used by the electrocardiogram. A continuous or alternated current is injected into the body, the voltage is measure, the ratio being the impedance. Several seconds are required to switch from the electrocardiogram detection to bio-impedance detection in order to allow transients in analog filters to complete.
2.4 Altri sensori
The device will include other sensors. One microphone to detect cardiac sound (phonocardiogram) and respiratory sounds. The signal is sampled a 4 kHz with 16 bits resolution. One CMOS photodiode to detect the light reflect by tissues under the skin. The two LEDs emit light in the red and infra-red band. The signal is used for the photoplethysmogram (oxygen saturation estimation).
3. INDEPEDENT PROCESSING
The software which is run on the microcontroller is designed to filter signals coming from each sensor. The first step in the processing chain does include an independent filter chain for each signal.
The goal of the independent processing is to remove the noise and detect artifacts. Out of band noise is removed by bandpass filtering. Cleaned signals are used to estimate significant parameters (heart rate, respiratory rate, etc). Artifacts (false signals generated by motion) and anomalous conditions (electrode disconnection) are detected, obviously in this case the parameters are not available.
Parameter extraction is prone to false negatives (pattern is present but is not detected) and false positive (pattern detected even if not present). A high sensitivity would result in the correct detection of all patterns present in the signal but would also result in an higher false positive rate.
In the case of biomedical applications, it is preferred to have a high precision, to be sure that when the pattern is declared detected it is really present into the signal. Unfortunately this results in a higher false negative rate.
3.1 Electrocardiogram signal processing
Two different processing chains were implemented and evaluated. The first one is the Hamilton-Tompkins [1][2] processing chain, used by a well-known publicly available algorithm for heart rate estimation (OSEA). The chain consists in the following blocks: a low-pass filter with a 16 Hz cut-off frequency that eliminates high-frequency noise (muscular noise and 50-60 Hz power-line interference), a highpass filter with a 8 Hz cut-off frequency which removes baseline wander, a block that computes the first-order derivative (as peaks corresponding to heart beats generally present considerable slopes), a 80 ms moving average filter which highlights peaks, and an adaptive threshold detection logic.
The second chain is based on the Pan-Tompkins algorithm, and it has been chosen for the BGW prototype. While the filtering chain is essentially the same, the slope is squared and not taken in absolute value, and the moving average is performed using a 150 ms window instead of a 80 ms one. This results in a more effective peak detection, even in the case of low-amplitude peaks, in spite of a greater computational effort.
Fig.3 Results of the electrocardiogram processing performer by the Pan-Tompkins chain and the simple adaptive-threshold logic for the detection of peaks related to heartbeats. Red: false positives (events detected but not present). Green: false negatives (undetected events which instead took place). Black: Correct classification but on the wrong part of the waveform (e.g. P wave instead of QRS complex).
The first experimental results show two critical situations: electrodes disconnection and noise overlapping the spectral bandwidth of the signal of interest. They both lead to an erroneous detection of R wave peaks, with a consequent wrong evaluation of the heart rate (Fig. 3). The algorithm was improved in three ways. First, an evaluation of the signal variance allows the detection of critical events (e.g. electrodes disconnection, high-frequency noise), because the variance of the signal during these events is significantly higher. The second improvement is based on the computation of the similarity (matching) with a reference waveform (template). The similarity is evaluated by detecting the local minimums of the sum of the absolute differences (SAD) between the current signal portion and the template [3]. Finally, a template database is used to make the algorithm independent of the acquired waveform - this is necessary because the prototype must function regardless of the electrodes placement.
Fig.4 Processing chain and detection logic for heartbeats (R wave peaks)
Templates are updated according to recognized patterns. This grants an adaptation to the specific ECG waveform of the patient. The detection logic is showed in Fig. 4, the template database is instead presented in Fig. 5 with the results of the processing (to be compared with Fig. 3 in order to understand the performance improvement).
Fig.5 (left) database of waveform templates used for the R-wave peak classification; (right) results from processing: in the upper plot, the classification performed by the standard reference algorithm (OSEA), which results in 10 false positives, 24 false negatives and the proper classification of 25/49 beats; in the lower plot the classification performed by the modified version of the algorithm (STM), which results in 0 false positives, 10 false negatives and the proper classification of 39/49 beats.
3.2 Accelerometer signal processing
The accelerometer signal is filtered and processed in order to extract a parameter corresponding to the total activity, counting steps and therefore measuring the physical activity level. The detection is performed by identifying steep transitions on the accelerometer axes (Fig. 6, on the left). During these events, the ECG signal is affected by noise (probably due to pectoral muscles balancing during the walk). A more sophisticated elaboration would permit the estimation of posture (prone, supine, standing), acts (walk, run) and anomalous events (falls).
Fig.6 Synchronised detection of electrocardiogram (upper waveform) and acceleration along three axis (x, y, z); (left) the patient is walking (noisy ECG); (center) patient standing still (heart rhythm present also on the xyz signals); (right) patient standind still but breathing heavily (breathing rhythm also on ECG and XYZ).
Heart and respiratory rates can be estimated from the accelerometer signal as well (Fig. 6, center and right). As a matter of fact, the accelerometer is capable of detecting weak signals with a frequency lower than 150 Hz, providing a good mechanomyography and a satisfying phonocardiogram. Unfortunately this detection is possible only when the patient is not moving. The respiratory rate can also be extracted from the ECG signal, by processing the R wave peaks envelope (Fig. 6, on the right).
A more sophisticated processing might allow the estimation of blood pressure from the acoustic/accelerometer signal [4].
3.3 Bio-impedance signal processing
The impedance signal acquired using a DC current allows the estimation of the electrical contact quality. The signal acquired using high-frequency currents (30-50 KHz) instead permits to extract the respiratory rate (however the electrodes optimal placement is different from the ECG one).
4. JOINT PROCESSING
After an independent processing, the prototype performs a joint processing.
4.1 Fusion of same parameters from different signals
The same parameter can be estimated from signals acquired using different sensors. A more reliable estimate can be obtained by combining signals from different channels. For example, the heart rate is primarily estimated using the electrocardiogram. The same signal can be confirmed by the results of the processing of the photoplethysmographic or the acoustic signal. When the patient is not moving, the accelerometer signal provides an useful estimate as well. Similarly the respiratory rate is detected mainly from the photoplethysmographic signal and from the impedance signal. The same signal can be confirmed by the microphone and, during stillness, by the accelerometer.
4.2 Comparison of different parameters from different signals
Since several parameters are available, it is possible to combine them to extract a more reliable information, or even a new one. For example, the photoplethysmographic signal is affected by noise artefacts that can be estimated with the accelerometer. The system performs a real-time cancellation of these artefacts, providing a more reliable information. In the same manner, the heart rate extracted from the electrocardiogram is compared with the estimate of the physical activity performed using the accelerometer, distinguishing between anomalous (stillness combined with tachycardia) and normal situations (running with a high heart rate).
5. PROTOCOL OF USE
The electronic patch is applied by paramedic personnel. In the start-up phase the correct detection of parameters is checked in different situations (e.g. stillness, walking or running). In this phase the software learns the typical waveforms and the typical values of the parameters of the specific situations. Therefore the device adapts to the specific patient. Afterwards the software performs a continuous monitoring, recording the parameters resulting from the processing. If anomalous events are recorded, the device records both the parameters and the raw signals, making them available for a subsequent examination. The software periodically acquires signals from other devices using the wireless connection (e.g. sent by an electronic scale) and sends the processing results to a central server.
6. BIOMEDICAL APPLICATIONS
The device has been conceived and designed to be used in a clinical context. However it is possibile to develop applications also in other fields: rehabilitation, training, healthcare.
7. BIBLIOGRAFIA
[1] B.-U. Kohler, C. Hennig and R. Orglmeister, "The principles of software QRS detection", IEEE Eng. Med. Biol. Mag., vol. 21, pp. 42-57, Jan./Feb. 2002.
[2] P. S. Hamilton and W. Tompkins, "Quantitative Investigation of QRS detection rules using the MIT/BIH arhythmia database", IEEE Trans Biomed. Eng., vol. BME-33, pp. 1157-1165, Dec. 1986.
[3] S.E. Dobbs, N. M. Schmitt, and H. S. Ozemek, "QRS detection by template matching using real-time correlation on a microcomputer", J. Clinic. Eng., vol 9, pp. 197-212, 1984.
[4] C. Tranulis, L. G. Durand, L. Senhadji and P. Pibarot, "Estimation of the Pulmonary Arterial Pressure by a Neural Network", Med. Biol. Eng. Comput. 03/2002, 40(2): 205-12.
--Andrea Vitali, Marco Pessione, STMicroelectronics
