1. The expert system toxicity detection algorithm was compared with the current algorithm using data gathered by the USACEHR research team during several years of development and testing of the system: The neural network training set included baseline (pre-exposure) data from multiple laboratory and field studies. Field studies included normal seasonal and diel variations in water quality parameters including pH, temperature, dissolved oxygen, and conductivity. The neural network system was tested using data from laboratory studies using varying concentrations of toxicants, with concentrations both above and below the 96-h LC50. Fish group responses of the neural network system were compared with the current statistical analysis system for the fish biomonitor (10 laboratory studies) Automated biomonitoring systems continuously monitor water quality and provide rapid notification of developing toxicity caused by a wide range of substances. An important goal for biomonitors is to maximize sensitivity to toxicants while minimizing false alarms that may be caused by non-harmful variations in water quality. Significant improvements in toxicity detection without an increase in false alarms are possible through the use of novel data processing and neural network modeling approaches developed for an automated fish biomonitoring system. Toxicity detection is based on simultaneous analysis of ventilatory and movement behavior of a group of eight fish (bluegill, Lepomis macrochirus) and water quality parameters (pH, dissolved oxygen, temperature, and conductivity). A general neural network model of fish behavior is used that does not need to be re-calibrated for each individual fish. The model can detect abnormal patterns in fish behavior associated with toxicity with a better signal-to-noise ratio than the present statistical approach, while distinguishing between changes in fish behavior due to toxicity or water quality variations. The automatic data interpretation is a part of Biomonitor Expert, a Windows-based program that addresses all aspects of the biomonitoring process, including data collection and storage, construction of neural network models of toxicity, user-friendly interfaces and remote notification tools. A modular design enables easy re-configuration of the system, inclusion of different data collection and processing schemes, and application to different biomonitoring applications and toxicity sensors. EXPERT SYSTEM FOR REAL-TIME BIOMONITORING OF ENVIRONMENTAL TOXICITY T.R. Shedd 1, D. Wroblewski 2, M.W. Widder 1, W.H. van der Schalie 1, J. Viloria 2, M. Green 2, R. Siegel 3, and M. Buller 3 1 U.S. Army Center for Environmental Health Research, 2 Intelligent Automation Corporation, 3 GEO-CENTERS INC BioMonitor Expert: Comprehensive Software Program for a Real-time Biomonitoring System Data collection and management, user interfaces, and real-time data interpretation employing advanced models of fish ventilatory behavior Toxicity detection based on anomaly detection and standard model of fish behavior Interfaces for automated process control Water sampling Water flow control Communications Web accessible Remote notification about toxicity status (alarm) Data visualization Graphs of historical data Analysis of historical data Statistics Neural network model Model updates Data labeling Retraining of the (neural network) fish behavior model MS Windows program Under development for the fish-based system To be completed in 2003 Disclaimer The views, opinions, and/or findings contained in this report should not be construed as official Department of the Army position, policy, or decision unless otherwise designated by other official documentation. Research was conducted in compliance with the Animal Welfare Act and other Federal statues and regulations relating to animals and other experiments involving animals, and adheres to principles stated in the Guide for the Care and Use of Laboratory Animals, National Academy Press, Washington, DC, 1996. Introduction Automated biomonitoring complements chemical testing Continuous monitoring (24/7) Short response time: early warning systems Non-specific response: reaction to a wide range of toxic conditions Sensitive to effects of interaction of many pollutants/toxins Best for detection of Developing toxicity Transient toxic events Unsuspected toxins Mixtures of toxins Different organisms may be used for monitoring of water toxicity Fish-based biosensors are used in the present system As a result of their neuromuscular activity, all fish generate a low-level electric field detected with non-contact electrodes. Fish ventilatory and locomotor behavior information that can be extracted from the electrical signal includes ventilatory rate and depth, cough rate, and whole body movement Changes in ventilatory and movement patterns can indicate both developing toxicity and variations in certain water quality parameters (e.g., temperature or dissolved oxygen) A Comparison of Ventilatory Data Analytical Methods Biomonitoring System Abstract Remote electrodes Signal amplification On-demand water samples for analytical measurements Online chemical sensors Drain Monitoring Chamber (8 fish) Electronic data Ventilatory Frequency Spike Cough Ventilatory Depth High Frequency Cough Cough Frequency Whole Body Movement Ventilatory and Movement Parameters Modular design enables easy reconfiguration of the system and application to different biomonitoring applications and toxicity sensors Web-enabled user interface provides access to the system configuration management and remote access to bio-monitoring results Current System Expert System Ventilatory parameters (individual fish) Manual determination of whether water quality changes caused response Collect pre- exposure data (96-h) for each fish Calculate confidence limits for each fish and parameter Determine if individual fish are outside confidence limits for any parameter Group response if more than 70% of fish respond Water quality data Materials and Methods Results Receiving Water Protection Source Water Protection Applications Laboratory Toxicity Screening New York City Fort Detrick Current System On-line Data Output Toxic Event 7 pH 8.24 Print Interval: 56 Data Collection Time: Fri Apr 04 19:07:18 1997 Control Groups Expert System Current System No detection Detection No detection Detection 9 0 1 0 Exposure Groups Expert System Current System No detection Detection No detection Detection 4 1 4 21 Current System Quicker Expert System Quicker Current System Quicker Expert System Quicker Baseline required for individual fish: The current system requires baseline data for each fish to determine normal limits for the ventilatory and movement parameters. Group fish responses based on summed individual responses without consideration of water quality parameters Water quality perturbations in baseline period may affect sensitivity of system response to toxicants Fixed, individual fish baseline periods are only valid for environmental conditions encountered during baseline period Must use expert judgment by system operator to separate group responses due to routine water quality variations from those due to toxicity Novel, time-dependent normalization procedure: The expert system uses a normalized model, independent of idiosyncratic characteristics of individual fish and does not require a baseline period to define normal behavior for each fish Neural network model of fish behavior and the water quality parameters: abnormal behavior is detected as ‘anomaly’ or, inability of the model to reproduce the measurement data for normal behavior Expert system incorporates fish responses to water quality perturbations in neural network model Distinguishes anomalies due to toxicity from those due to routine variations in water quality parameters (e.g., temperature or dissolved oxygen) Learning system: as the system matures and more data becomes available, the neural network model is retrained to include the newly discovered patterns Conclusions Both the current and expert system approach detect emerging toxicity before lethality The expert system approach detects developing toxicity more rapidly than the current system, in most cases In some cases, the expert system responded at lower toxicant concentrations than the current system The expert system distinguishes behavioral changes due to variations in water quality parameters from toxicant-related changes Further research is needed to fully optimize the detection algorithms New signal processing approaches for characterization of the bioelectric signals Refinements in toxicity detection algorithm