Innovative DSP architectures are fostering the development of advanced medical diagnostic and monitoring equipment that simply was not possible even a few years ago.
The Diopsis dual-processor, floating-point VLIW DSP from Atmel Corp. is a case in point. Native support of complex arithmetic, 40-bit extended precision and a VLIW architecture that supports 1 Gflops at 100 MHz allow the execution of advanced adaptive algorithms that Canadian Medical Technologies Inc. (Cana-met) uses in its portable, noninvasive diagnostic and monitoring equipment. The processor's low clock rate and SoC integration of an ARM7 and peripherals help conserve power and space.
The 7 In One integrated vital signs monitor and Piesometer MK-1 blood pressure monitor from Canamet both use the Diopsis as the system controller and DSP.
Until recently, there were just two ways to monitor a patient's blood pressure. Using the auscultatory method, a medical practitioner with a stethoscope and mercury sphygmomanometer monitors the Korotkoff sounds caused by blood flowing through the brachial artery. The other method, employed by automatic blood pressure machines, relies on an oscillometric technique that estimates blood pressure based on tiny pressure fluctuations caused by the Korotkoff sounds and detected by the same pressure transducer that monitors the pressure in the pressure cuff. In vibration-intensive environments, the vibration itself can cause pressure fluctuations that reduce measurement accuracy.
Canamet has overcome those issues by using adaptive signal-processing techniques that include adaptive interference cancellation, bandpass filtering and peak discrimination algorithms.
Canamet's Piesometer MK-1 portable blood pressure monitor has a primary acoustic sensor (placed on front of the patient's arm) that collects the Korotkoff sounds and any environmental noise. A secondary acoustic sensor (on the back of the arm) collects only the environmental noise. The adaptive interference cancellation algorithm is a nonlinear filter that removes the interference measured by the secondary sensor from the desired signal received by the primary sensor that was corrupted by environmental noise. The bandpass filter removes noise outside the frequency of interest. Finally, the peak discrimination algorithm extracts valid peaks from the Korotkoff sounds in the acoustic signal that result from heartbeats.
The first step is to isolate peaks that are greater than the noise floor level that is determined by the peak discriminator. The peaks are then further examined to determine whether they satisfy some periodicity and constancy in repetition (i.e., possible heartbeats), as would be expected from a human heart. Peaks that do not satisfy those constraints are discarded; however, a degree of arrhythmia is accounted for during this process. The result of the peak discriminator is a series of constantly repeating periodic peaks. This process also eliminates random peaks that may be due to strong transient noise effects.
The results are used to estimate the patient's pulse rate and derive the patient's blood pressure. The systolic blood pressure is defined as the point at which the positive slope of the envelope of the sequence of detected peaks is greatest. The diastolic pressure is at the point of greatest negative slope of the envelope. The times that these two events occur are referenced to the data acquired by the pressure transducer, which provides a time-varying measurement of the pressure in the deflating pressure cuff. The Piesometer MK-1 uses the auscultatory technique and has proved to have accuracy equivalent to that of a medical practitioner.
Advanced adaptive algorithms are computationally intensive, requiring the execution of billions of operations per second (Gops) and relying heavily on the use of fast Fourier transforms for such frequency-domain operations as time delays and spectral analysis. Conventional DSP architectures are not up to the task.
Fixed-point computing architectures do not possess the appropriate processing capabilities to execute these computationally intensive algorithms efficiently, and they tend to introduce inaccuracies in the weight vector calculation that actually increase the noise in the system. Floating-point arithmetic enables much more accurate calculations and provides faster development cycles because the C code does not have to be translated to a fixed-point format.
Innovations in the Diopsis DSP were indispensable to the development of Canamet's products. The 40-bit extended precision permitted extremely accurate computation of the adaptive weights in the adaptive interference canceller algorithm. While conventional floating-point DSP architectures typically require two cycles to execute an FFT, the Diopsis instruction set provides native support for the complex domain arithmetic required for single-cycle FFT execution for frequency-domain operations.
The very long instruction word, large register files and efficient architecture in the processor block let Diopsis execute 15 operations per second twice the throughput per cycle of any other DSP. Thus, while most DSPs must have a clock of 250 MHz to achieve the required Gops throughput, Diopsis can achieve 1.5 Gops, or 1 Gflops, with a clock rate of only 100 MHz. This keeps power consumption to only 750 milliwatts.
Finally, integration of the DSP with the ARM7 MCU, eight A/D and D/A interfaces and other peripherals provides a truly single-chip solution.
Stergios Stergiopoulos (firstname.lastname@example.org) is president and Amar Dhanantwari (email@example.com) is chief scientist at Canamet Inc. (Toronto).
See related chart