音声ブラウザご使用の方向け: SKIP NAVI GOTO NAVI

CONTROL OF NEURAL PROSTHESES I: EVENT DETECTION USING MACHINE LEARNING

Richard Williamson, Brian Andrews, Raymond Au Department of Biomedical Engineering, University of Alberta,Edmonton, Canada

ABSTRACT

Machine Learning has been explored as a control method for FES. Modern, micro-machined accelerometers designed for use with airbag control have provided a potential sensor that can be implanted or conveniently mounted on an orthotic brace or belt without the neccesity for precise anatomical alignment. Rough Sets has opened doors for the speed of rule implementation in control systems as well as noise tolerance seen in fuzzy computing methods. Rough sets has shown that it can improve classification of accelerometer pattern in comparison to ALN. The technique of classifying accelerometer pattern with Rough Sets will allow for new control paths in both implant and surface FES.

BACKGROUND

Rule base control of FES has been explored by many. EMG triggers [1],were considered as natural feedback signals for FES. External triggers based on force and angles [2] or accelerometers [3] have been combined with hand-crafted rules for FES control. Improvements can be made on hand crafted rules by fuzzification and machine learning [4]. Supervised machine learning of rules has also been explored. Rule induction [5], neural networks [6] have been explored as possible FES control methods. Adaptive Learning Nets [7] have shown themselves to be capable of learning a pattern for FES required in a case of hemiplegic gait. Recently, Rough Sets have been shown to be able to discriminate gait phases in normal walking [8]. A comparison between these two methods of learning will provide a basis for continuation of supervised learning methods for control of FES. An Adaptive Learning Network (ALN) [9], composed of linear threshold units(LTU), AND gates, and OR gates, is a method of piece wise linear fitting a desired output curve from an input pattern. Linear threshold fit data within a particular region of input space and are combined using the AND and OR gates. A decision tree is constructed of the LTU, AND and OR gates. Rough Sets is purely a classification method. The input space, defined by the attributes, is partitioned into regions in which a minimum percentage of the space is within one output class. Regions of an output class need not only contain examples of that class. This allows for greater rule generalizations and noise tolerance than other rule induction methods.

RESEARCH QUESTIONS

1) Are ALN and/or Rough Sets capable of determining gait phases as from accelerometer data? 2) Which accelerometer attributes are important for gait phases determination?

METHOD

Three +/-5g ADXL05G accelerometers were mounted on a each calf brace. The accelerometers were mounted on the mounting boards provided with the accelerometers and were prepared to have a range of +/- 2 g and 1st order filter at 33 Hz. The accelerometers were sampled using 12 bit A/D at 100 Hz. The arrangement of the accelerometers was not orthogonal, but formed a three dimensional basis. Interlink force sensing resistors were placed in the insole of the subject's shoe under the heel, lateral metatarsal and medial metatarsal. Gait phases, as described by Perry [10], were determined superficially from the FSRs using the following rules. [Table 1] Experimental data was taken for walking of 5 able-bodied individuals (table 2) around an oval and figure 8 pattern. The oval pattern was approximately 24 m; the figure 8 pattern was approximately 26 m. [Table 2]

RESULTS

For each trial a confusion matrix of classification rates was evaluated. The confusion matrix is simply a table that counts what the predicted state is during a sample. A diagonal element of the confusion matrix counts the number of correct predictions of a state. A filter was designed to improve the performance of the classification. It used three future points, enforcing a 3 ms delay in the processing of the data. This delay was not accounted for in processing of data, i.e. the filter results are counted to occur at the time of the sample, the time sample being filtered is not the furthest ahead in time. The filtered confusion matrix diagonals are provided as percentage correct classifications. The ATREE3.0 program was tested using training files of 15 s with input attributes of the present accelerometer sample, it first past point, and a simple difference of these two values. The training files for the DataLogicR 1.5 program contained the simple difference and two past points. The following table displays the training and testing classification rates for filtered and nonfiltered ATREE3.0 and DataLogicR 1.5 output. [Table 3] Rough Sets was also tested on a walking pattern combining figure 8 and oval data. Training files were generated containing 10 seconds of data from each the oval and the figure 8 for one subject. Rules were tested on the remaining data. [Table 4] For practical uses, tests were also run on the rough sets program to determine the importance of the past points and simple difference attributes on the classification rates as well as the importance of each accelerometer. [Table 5] [Table 6]

DISCUSSION

From the above results, DataLogicR 1.5 has shown its ability to classify gait phases from accelerometer signals. DataLogicR 1.5 has also shown a better ability to classify this accelerometer data into gait phases by approximately 10%. DataLogicR 1.5 has eliminated the a disadvantage of inductive learning in comparison to ATREE3.0 by being more noise resilient than previous induction methods. DataLogicR 1.5 can capably classify data using only a simple difference attribute if three accelerometers are used, and the anterior accelerometer is the most important for phases classification. Improvements in using both DataLogicR 1.5 and ATREE3.0 and the application of heuristic knowledge should provided for better classification results in the future. As the ALN has previously been shown as a potential controller of FES in hemiplegic gait employing FSR as input signals, accelerometers placed on the shank of the leg and employing Rough Sets as a learning method has control possibilities. Rough Sets has also shown an ability to classify over a more general gait pattern seen in the combination of the figure eight and oval. Error rates of ~20% in generalization seem adequate in an primary trial.

References:

[1] J.Symons, D.McNeal, R.Waters, J.Perry, Trigger Switches for Implantable Gait Stimulation, RESNA 9th Ann. Conf. 319-321 (1986)

[2] B.Andrews, R.Barnett, G.Phillips, C.Kirkwood, N.Donaldson, D.Rushton, T.Perkins, Rule-Base control of a hybrid FES orthosis for assisting paraplegic locomotion, Automedica,11:175-199 (1989)

[3] A. Willemsen PhD. Th, Accelerometers for Gait Assessment in Functional Neuromuscular Stimulation

[4] S.Ng, H.Chizeck, A fuzzy logic gait event detector for FES using Firmware Transitional Logic,Proc. IEEE EMBS 10th Ann.Conf.1562-1563

[5] C.Kirkwood, B.Andrews,Finite-state control of FES systems:Application of AI inductive learning techniques, Proc. 11th Ann. IEEE EMBS Conf. 1020-1021 (1989)

[6] B.Heller, P.Veltink, N.Rijkhoff, W.Rutten, B.Andrews,Reconstructing muscle activation during normal walking: A comparison of symbolic and connectionist machine learning techniques, Biol. Cyber. 69:327-335 (1993)

[7] A.Kostov, B. Andrews, D. Popovic, R. Stein, W. Armstrong, Machine Learning in Control of Functional Electrical Stimulation Systems for Locomotion, IEEE Trans. Biom. Eng. 42:6:541-551 (1995)

[8] B.Andrews, R.Au, R.Williamson, Event Detection for FES Control Using Rough Sets & Accelerometers, Proc. 2nd Int. FES Symp., 187-193, (1995)

[9] Armstrong, W.W., Thomas, M. M., ATREE release 3.0 software, Dendronics Decisions Ltd. 1995

[10] Perry, J., Gait Analysis: normal and pathological function, Thorofare, N.J. : Slack inc., 1992

Acknowledgements:

Richard Williamson received a graduate scholorship from NSERC (National Research Engineering and Research Council);Raymond Au received a student assitanship from AHFMR (Alberta Heritage Foundation for Medical Research). Address Richard Williamson 10th Floor Clinical Science Building University of Alberta Edmonton, Alberta T6G 2J5 email:Richard.Williamson@ualberta.ca