Analyss of EEG of shooters Nuno Bandera Computer Scence Dept. New Unversty of Lsbon Qunta da Torre 2825-114 Caparca Vctor Lobo Portuguese Naval Academy, Escola Naval, Alfete, 28 ALMADA Fernando Moura-Pres Computer Scence Dept. New Unversty of Lsbon Qunta da Torre 2825-114 Caparca Abstract Electroencephalogram (EEG) data was collected from competton-level shooters durng practce sessons and searched for relatons between ths EEG data and the shooter s score on target. To ths purpose Self-Organzng Maps (SOM), cross-correlaton analyss and scatter matrces based technques were appled to dentfy whch varables were relevant for the correct classfcaton of EEG data vectors nto the correspondng target score classes. Keywords EEG, SOM, feature selecton, PCA, Scatter matrces 1. Introducton Accordng to competton-level shootng experts, the shooter s concentraton on the shootng task s a most relevant factor of success. Wth ths assumpton n mnd we collected electroencephalogram (EEG) data from competton-level shooters durng practce sessons and tred to fnd relatons between the EEG data and the shooter s score on target. To ths purpose we used Self-Organzng Maps (SOM), crosscorrelaton analyss and scatter matrces technques to dentfy whch varables or sets of varables were relevant for the correct classfcaton of EEG data vectors nto the correspondng target score classes. 2. Data acquston The EEG recordngs were made usng a Brantroncs ISO-132/Control-132 system connected to a Data Translaton (DT-2839) sgnal acquston board, nstalled n a PC runnng Wndows95 and our own real-tme EEG acquston software. The electrodes were postoned on the shooter s head mmedately pror to the practcng sesson (accordng to the standard 1-2 electrode placng scheme [2][1]) and bookmarks were placed on the EEG tme seres accordng to each shot. Due to the muscular actvty of the shooter between shots (body and arm postonng, relaxng, browsng of target score, etc) most of the contnuously recorded EEG data was useless. By vsual nspecton of the 21 EEG channels tme seres we observed that the longest artfact-free EEG sequences avalable pror to any vald shot were 5 seconds long. We then appled a wndowed Fast Fourer Transform to each EEG channel s extracted 5 seconds seres obtanng the d spectra wth approxmately 1Hz resoluton between 1Hz and 3Hz. Snce we had to remove 2 EEG channels due to the extreme presence of nose and artfacts and the 2 ears electrodes, we were left wth 17 channels of 3 frequency bns each leadng to a total of 51 features.
3. Self-Organzng Maps Self-Organzng Maps (SOMs) were used as a tool for projecton of hgh dmensonal data nto human-understandable 2D maps. Our objectve was that of usng ths 2D projecton to vsually explore and dentfy whch subset of features was sutable for an effcent classfcaton. If we could dentfy a reasonably defned set of clusters on a SOM then we would have reasons to beleve that we had found a feature subset [3]. SOMs are a very well documented exploratory technque [4][5] so we ll provde only a bref descrpton of ts nner workngs: 1. For a gven tranng pattern x: 1.1 - Calculate the dstance of each neuron to the tranng pattern x 1.2 - Fnd the neuron wth smaller dstance, and call t the wnner W 1.3 - Change the network neurons wth a functon G, whch depends on the learnng rate α, the dstance d to W (n the output plane), and the neghborhood functon F. Due to the nature of the neghborhood functon, only the neurons closer to W (n the output space) wll be changed. (Update phase) Update the learnng rate α and the neghborhood functon F accordng to some rule Repeat steps 1 and 2 for the next tranng pattern, untl some stoppng crtera s reached. Some nterestng results were found by the use of SOMs but the strongest conclusons that we were able to draw from ths exploratory analyss was: results could be acheved only f we looked at our data from a shooter-orented perspectve and appled a somehow more effcent feature subset selecton method. Fgure 2 shows one of the best maps that could be obtaned for all shooters smultaneously. Some clusters can be dentfed but there are no clear fronters and there s stll a sgnfcant level of dsperson. Fgure 1 - Legend for the map Fgure 2 - EEG channel T3, all shooters 4. Statstcal technques used We tred two dfferent approaches to feature selecton usng standard statstcal descrptve technques: cross-correlaton and scatter matrces. 4.1 Cross-correlaton Features were selected usng crosscorrelaton accordng to the followng crtera: 1. Hghest cross-correlaton wth the shooter s score 2. Lowest cross-correlaton between the selected features Our purpose was that of fndng mutually uncorrelated features that could both explan the shootng score. Ths should lead to an occupaton of most of the subspace defned by the subset of selected features therefore enhancng the possbltes of defnng lnear dscrmnant surfaces. 4.2 Scatter matrces Scatter matrces are commonly used n dscrmnant analyss to formulate crtera of class separablty [6]. Wthn-class scatter matrces are used to evaluate the scatter of the data vectors around the classes expected vectors and for L classes are defned as:
where Σ s the covarance matrx for class s data vectors. To evaluate the scatter of the classes expected vectors around the expected vector of full dataset, a betweenclass scatter matrx s gven by b = where µ s the dataset s expected vector. Fnally the mxture scatter matrx s the well-known covarance matrx of the full dataset: m S These matrces can be used to defne several dfferent class separablty crtera. We chose to use where tr(m) s the trace of the matrx M. We then maxmzed ths crteron to select the best subset of features. Ths technque s computatonally very heavy due to the fact that J must be evaluated for every possble subset of a gven number k of features. We tred to solve ths combnatoral exploson problem by mplementng branch-and-bound enumeraton [6], n order to guarantee the global maxmum wthout havng to explctly evaluate every possble subset of features. So far we had only lmted success wth ths technque, producng subsets of two features as we shall present n the next secton. 5. Results L S w = P Σ L = 1 P ( ì = 1 ì We were nterested n applyng these technques both to our orgnal data and to the prncpal components as obtaned by standard prncpal component analyss (PCA). PCA was also a very nterestng approach for us because we wanted to search ahead for the reasons that lead the SOMs to such dsperson. By usng PCA we could know f t was due to correlaton between our ntal features or f we smply dd not have o )( ì ì T [( X ì )( X ì ) ] = Sw Sb S å + = J tr 1 ( S = S w b) o ) T enough data shooter ndependent analyss. Snce we had more features than data vectors we had to splt our full set of features n half (accordng to the left/rght bran hemsphere) n order to be able to apply PCA. To such purpose three dfferent datasets were assembled: DS1 Left bran hemsphere DS2 Rght bran hemsphere DS3 Full dataset Frst we wll present the results obtaned usng the full dataset (DS3) and then our results usng left/rght hemsphere prncpal components datasets (DS1/DS2). 5.1 Full dataset (DS3) Ths dataset was parttoned n sx dfferent datasets, one per shooter. We then ran both feature subset selecton crtera on each dataset and the results are presented n Table 1 wth each feature represented by ts EEG channel name and frequency bn (n brackets). Cross correlaton Scatter matrces J 1 F7(6), T5(16) F7(6), T5(16) 2 O2(21), T3(1) F7(25), T3(1) 3 F8(1), Fz(1) F4(14), Fz(3) 4 C3(15), F3(29) P3(3), F8(26) 5 P3(15), T4(21) Fp1(5), F8(5) 6 F3(15), F3(19) P4(1), P4(2) Table 1 Pars of features selected from DS3 In all cases but one our choce of features was dfferent from that obtaned usng the scatter matrces crteron J. Furthermore, usng cross-correlaton we were able to fnd pars of features more sutable for pecewse lnear separaton of shootng scores classes. In some cases only mnor mprovements were acheved, as s the case of shooter 2 (see fgures 3 and 4), but n others we were able to mprove class separablty by a large margn. Such s the case for shooter 6 (see fgures 5 and 6).
4 3 2 3 25 2 15 1 1 5 4 3 2 1 2 4 6 8 Fgure 3 - Shooter 2, F7(25) x T3(1) 5 1 15 2 25 3 6 x 14 5 4 3 2 1 Fgure 4 - Shooter 2, O2(21) x T3(1) 5 1 15 Fgure 3 - Shooter 6, P4(1) x P4(2) 1 2 3 4 Fgure 6 - Shooter 6, F3(15) x F3(19) In fgure 5 we can see that the scatter matrces crteron leads to a coherent result: low wthn-class scatter and a reasonable dstance from each classes expected vector to the dataset s expected vector. Nevertheless ths s a bad soluton because the classes varance leads to overlappng, therefore preventng us from fndng a pecewse lnear separataton. A much better subset of features from the pecewse lnear separablty perspectve would be the soluton found usng the cross-correlaton crtera (shown n fgure 6). 5.2 Prncpal components of DS1/DS2 After performng PCA on DS1 and DS2 we chose two subsets of prncpal components. One where each component explaned at least 1% of the global varance and another where each component explaned at least.1% of the global varance. The results were: DS1 DS2 > 1 % 16 (82.8%) 15 (82.4%) >.1 % 57 (96.%) 57 (96.%) Table 2 Number of prncpal components
Table 2 also shows (wthn brackets) how much of the global varance was explaned by each selected subset of features. The scatter matrces crteron J nvarably selected components 1 and 2 when appled to any of these datasets. A plot of the shootng scores usng ths two components can be seen n the followng fgure: 2 bad 1-1 -2-3 -2 2 4 6 8 Fgure 7 - DS2's prncpal components 1 and 2 The results for the left hemsphere where bascally the same and support the clam that nsuffcent data was avalable for buldng a shooter ndependent classfcaton system. The cross-correlaton analyss resulted n: DS1 DS2 >1% >.1% >1% >.1% PC 1 1 1 1 1 PC 2 2 45 2 2 PC 3 14 24 13 13 PC 4 12 2 9 33 PC 5 8 25 3 25 Table 3 Frst fve prncpal components selected by cross-correlaton As can clearly be seen n table 3, n most cases the frst two selected features were also components 1 and 2. The cross-correlaton analyss here was mostly useful for supportng the scatter matrces crteron n renforcng the concluson that a shooterndependent system wasn t attanable. 6. Conclusons any conclusons can be drawn. Nevertheless our man goal of fndng subsets of features that would lead to an effcent classfcaton of the shooter s scores accordng to the EEG data was successfully acheved. These results not only tell us that t s possble to buld such a system, but allow us to buld t wth an adequate level of generalzaton ablty, therefore avodng undesred overfttng ptfalls. Furthermore, the dentfcaton of effcent subsets of features was also mportant from an expert knowledge pont of vew because we were able to dentfy whch EEG channels and frequency ranges were relevant for dfferentatng between states that lead to hgher or lower shooter performance. 7. References [1] Marc Nuwer, Detrch Lehmann, Fernando Lopes da Slva, Shgeak Matsuoka, Wllam Sutherlng, Jean- Franços Vbert. IFCN gudelnes for topographc and frequency analyss of EEGs and Eps. Report of an IFCN commttee. Electroencephalography and clncal Neurophsology, 91 (1994), 1 5. [2] Una revson del sstema nternaconal dez-vente de colocacon de electrodos. Grass medcal nstruments. [3] Nuno Bandera, Vctor Lobo, Fernando Moura-Pres. EEG/ECG data fuson usng Self-Organsng Maps. Proceedngs of EuroFuson99. Jolver Press, 1999. [4] Tuevo Kohonen. Self-Organzng Maps. Sprnger-Verlag, 1995. [5] http://www.cs.hut.f/research/somresearch/ [6] Kenosuke Fukunaga. Introducton to Statstcal Pattern Recognton, 2 nd edton. Academc Press, 199. Our approach s essentally descrptve, shooter dependent and the whole process has to be appled to every new shooter before