Michael Factor

Physiologic trend detection and artifact rejection: a parallel implementation of a multi-state Kalman filtering algorithm

Using a parallel implementation of the multi-state Kalman filtering algorithm, we have developed an accurate method of reliably detecting and identifying trends, abrupt changes, and artifacts from multiple physiologic data streams in real-time. The Kalman filter algorithm was implemented within an innovative software architecture for parallel computation: a parallel process trellis. Examples, processed in real-time, of both simulated and actual data serve to illustrate the potential value of the Kalman filter as a tool in physiologic monitoring.

Real-time data fusion in the intensive care unit

A description is given of the process trellis, a domain- and hardware-independent software architecture. Its usefulness in building the Intelligent Cardiovascular Monitor, a real-time clinical decision-support system whose paving-stone interface gives a clinician an instantaneous overview of a patients status, is demonstrated. Issues of parallelism, real-time operation, and visualization are discussed.<<ETX>>

A parallel software architecture for building intelligent medical monitors

Intensive care units become more complicated each day as the number of devices developed to monitor various aspects of a patients status continues to increase. Intelligent monitors attempt to reduce this complexity by interpreting the data and presenting a high level summary to the clinician. We propose an innovative parallel software architecture for constructing intelligent medical monitors: theprocess trellis. The process trellis is an explicitly parallel structure, and therefore can take advantage of the performance gains available from parallel computing hardware. It does not, however, presuppose any expertise in parallel programming on the part of the application programmer. A prototype cardiovascular monitor has been built using this parallel software architecture. Preliminary testing of the monitor has shown that real-time cardiovascular monitoring, including data calculations, symbolic classification, and interpretation can be accomplished in real-time.

The process trellis architecture for real-time monitors

The process trellis is a parallel software architecture for building heuristic real-time monitors. These programs, for example Intelligent Cardiovascular Monitors, must process massive quantities of data in real time. It is natural to turn to parallelism to meet these computational requirements. The process trellis software architecture is intended to simplify the creation and maintenance of heuristic real-time monitors. To do this it must be 1) modular, 2) efficient and 3) predictable.nThis paper presents the goals and an overview of the process trellis. We have implemented a process trellis shell, which we describe. It is in use as the frame for an Intelligent Cardiovascular Monitor (ICM) which we are building with colleagues from the Yale School of Medicine. An analytical model is able to produce an upper bound on the time required by arbitrary trellis programs. Finally, we report on the predicted and actual performance of the ICM and synthetic programs consisting of roughly 7000 processes.

BIO-SPEAD: a parallel computing environment to accelerate development of biologic signal processing algorithms

We have created BIO-SPEAD (pronounced speed), a BIOlogical Signal Processing Environment for Algorithm Development. BIO-SPEAD is designed to accelerate development of complex algorithms which integrate information derived from single or multiple physiologic waveforms. BIO-SPEAD currently performs all of the basic analyses of several arterial blood pressure waveforms, and allows the user to utilize the results of those low-level analyses for development of more complex algorithms. We utilized a parallel programming architecture called the Process Trellis which keeps the different tasks, or processes, within BIO-SPEAD independent of each other. Additionally, we have developed a graphics interface to enable the user to visualize the waveform under analysis, the low-level system analysis, and the internal workings of the algorithm under development. The system has been used for several algorithm development projects and has demonstrated its utility.

The process trellis: a software architecture for intelligent monitors

The process trellis software architecture aids the programmer in building intelligent monitors for hierarchical domains by providing a way to guarantee real-time performance, structuring the interactions between modules, guiding the decomposition of the computation, and providing a well-defined interface with the external world. There is a natural parallel execution strategy for these decision processes which enables programs that would otherwise be unable to meet their real-time constraints to do so and permits speedup of programs without real-time constraints. One way to demonstrate a software architectures utility is by an existence proof; HEARSAY-II is such a proof for blackboard architectures. While still a research project, a process trellis shell, which implements most of the features of the process trellis software architecture, is being used for an intelligent cardiovascular monitor that is intended for eventual use in a cardiac intensive care unit.<<ETX>>

Parallel Multi-channel Biologic Signal Processing

We are building a prototype intelligent cardiovascular monitor (ICM) using the process trellis software architecture for parallel, real-time monitors. This paper describes our experience with the signal processing portion of the ICM. We implemented two versions of the multi-state Kalman filtering algorithm for simultaneous, multi-channel, biologic signal processing. One version had a separate process for each signal. This paper concentrates on a version which parallelized the analysis of a single signal. Unfortunately, unresolvable data dependencies within the Kalman filtering algorithm severely limited the performance gains. The process trellis is a software framework for combining heterogeneous, black-box processes (written in C) into a realtime monitor that executes on a parallel computer [6]. The process trellis itself is implemented in the machine-independent parallel programming language C-Linda@ [7].

The trellis architecture for intelligent monitors

The process trellis is a software architecture that imposes a simple and regular organization on a complex, diverse set of parallel processes. The organization it imposes is well suited to parallel real-time monitors and expert systems. The trellis architecture is based on a hierarchical graph of decision processes, reflecting the conceptual hierarchy of the problem domain. All processes in the graph execute continuously and concurrently, and they communicate among themselves and with the outside world using a simple and uniform protocol. The trellis is amenable to parallel real-time scheduling strategies, to abstraction strategies that synopsize the structure of complex trellises, and to visualization; it is a highly flexible architecture that seems well suited to a variety of domains. Extensive work on the trellis to date has focused both on design and analytical studies, and on the implementation, using C-Linda, of a substantial prototype trellis for patient monitoring in intensive care units.<<ETX>>

Chapter 15 - Process Trellis and FGP: Software Architectures for Data Filtering and Mining

Software tools for parallel programming used to be a chaotic area, with dozens of contenders promoting different models and methodologies. In the last few years, the field has settled down considerably: there are perhaps half a dozen approaches that are still “in contention.” These include message passing systems of various kinds, data parallel languages for synchronous architectures, explicitly-parallel Fortran variants—and (among a few others) Linda[7], which is the system weve used. C-Linda ® (the computing language C mated with the coordination language Linda) has been used to date mainly in scientific applications and in graphics (production Linda applications exist in areas like computational fluid dynamics, molecular modeling, radar cross sectioning, seismic simulations, genetic database search, genetic linkage analysis, ray-tracing and others). In our own group, however, weve long been interested in AI applications. The Linda model, based on a shared, associative, object memory, through which processes communicate “anonymously,” is appropriate (we believe) to a wide range of significant AI applications. In this article we discuss software architectures for data filtering and mining built using Linda.