- Email: [email protected]
Artificial neural network: A new diagnostic posturographic tool for disorders of stance
Clinical Neurophysiology 117 (2006) 1692–1698 www.elsevier.com/locate/clinph
Artificial neural network: A new diagnostic posturographic tool for disorders of stance Siegbert Krafczyk a,*, Simon Tietze a, Walter Swoboda a, Peter Valkovicˇ a,b, Thomas Brandt a a b
Department of Neurology, University of Munich, Marchioninistraße 15, D-81377 Munich, Germany 2nd Department of Neurology, School of Medicine, Comenius University, Bratislava, Slovak Republic Accepted 27 April 2006
Abstract Objective: To determine the accuracy of diagnoses made with artificial neural network techniques (ANNW) that identify postural sway patterns typical for balance disorders. Methods: Body sway was measured by means of posturography during 10 test conditions of increasing difficulty. From a database of 676 subjects 60 training cases (TCs) and 60 validation cases (VCs) were selected in which the following diagnoses had been established clinically: normal subject (NS), postural phobic vertigo (PPV), anterior lobe cerebellar atrophy (CA), primary orthostatic tremor (OT), and acute unilateral vestibular neuritis (VN). A standard 3-layer feed-forward ANNW, using the backpropagation algorithm, was trained with TCs, validated with VCs, and its accuracy tested on 5 new cases. Results: ANNW differentiated the established diagnoses with an overall sensitivity and specificity of 0.93. Sensitivity and specificity were 1 for NS and OT; for PPV, 0.87 and 0.96; for CA, 1 and 0.98; and for VN, 0.8 and 0.98, respectively. New subjects were identified with ANNW output variables of the true diagnoses between 0.73 and 1. Conclusions: ANNW differentiates postural sway patterns of several distinct clinical balance disorders with high sensitivity and specificity. Once designed and tested ANNW could be considered a black box, which each examiner can apply to predict a specific diagnosis even without a clinical examination. Significance: A promising diagnostic tool for disorders of upright stance in selected neurological disorders. q 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. Keywords: Upright stance; Posturography; Artificial neural network; Neurological diseases; Back-propagation-algorithm
1. Introduction Static posturography is used to analyze the body sway of unsteady patients with various neurological disorders. Until now clinicians have accepted this method as a tool for following-up balance disorders but not for reliably establishing a specific diagnosis. Their main reason has been that there are no specific analysis criteria of pathological sway patterns. So far diagnoses of only 3 * Corresponding author. Address: Department of Neurology, Klinikum Grosshadern, Ludwig-Maximilians-University, Marchioninistrasse 15, D-81377 Munich, Germany. Tel.: C49 89 7095 4803; fax: C49 89 7095 4801. E-mail address: [email protected] (S. Krafczyk).
conditions have been established on the basis of a routine analysis of body sway: (i) the 3 Hz sway in anterior lobe cerebellar atrophy (Dichgans et al., 1976; Diener et al., 1984); (ii) increased sway activity in the higher power spectra frequency band with a typical peak between 12 and 19 Hz in primary orthostatic tremor patients (Bronstein and Guerraz, 1999; Yarrow et al., 2001); and (iii) increased sway activity in the 3.5–8 Hz frequency band in patients with somatoform phobic postural vertigo (PPV) (Krafczyk et al., 1999). It is much more difficult to determine posturographic criteria for vestibular disorders like vestibular neuritis. The body sway of these patients increases in darkness during the first days after disease onset; however, because of the large variation in balance performance it cannot be differentiated from other balance disorders by
1388-2457/$30.00 q 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.clinph.2006.04.022
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
body sway alone (Strupp and Brandt, 1999; Strupp et al., 1998). Basically, however, it is possible to use a single or a combination of multiple parameters to characterize a certain balance disorder. In the following study we improved the discriminatory power of posturography by using multiple parameters obtained from the original data. These included sway path (SP) (for details see Hufschmidt et al., 1980); root mean square (RMS) (Brandt et al., 1981); and the sum activity of body sway in different frequency ranges after Fourier transformation analysis (FFT). Our goal was to determine the typical patterns of pathological sway for 4 disorders of upright stance: anterior lobe cerebellar atrophy, PPV, primary orthostatic tremor, and acute vestibular neuritis. Since, this approach increases the complexity of the data comparison to such an extent that it is impractical for routine clinical use, an artificial neural network (ANNW), as described by Duda et al. (2001), was applied for computational analysis of the sway patterns in the patients and in normal controls. The advantage of ANNW is that all posturographic parameters, as well as many pre-selected sway patterns, which have been identified by experienced examiners, can be included. The effectiveness of an ANNW model was newly shown by classifying the risk of falls in the elderly on the basis of an analysis of balance control during gait (Hahn and Chou, 2005). We recorded the postural sway characteristics of 4 neurological and vestibular disorders of upright stance and fed an ANNW with multiple parameters extracted from the raw data. Initially, the ANNW was trained with the sway parameters of 60 training cases (TCs) and was then validated with 60 validation cases (VCs). The VCs included 5 categories (4 balance disorders and normal sway); the diagnoses had been previously established clinically. The goal was to test the validity of the ANNW when used to determine the probable diagnosis of a balance disorder by comparing the data of new patients with the stored experience of the ANNW but without any access to further information.
2. Material and Methods 2.1. Patients A database of 676 adult patients, who had been assessed according to the same protocol and had undergone static posturography between 1998 and 2005, served as the basis for selecting 421 patients in the categories under investigation. The TCs and VCs (normal subjects included) were selected after the identification of congruent sway patterns with the help of confidence plots (see Section 2.3) and the verification of the diagnosis by clinical methods on the basis of the patients’ data sheet after discharge from the clinic and the establishment of the final diagnosis. The following
1693
groups were selected for the pattern classification of posturographic sway: 1. normal subjects (total number in the database, nZ41; 15 TCs and 15 VCs were selected, age:18–54y; mean: 29y; f:11; m:19). 2. Patients with the presumptive diagnosis of PPV (nZ225; 15 TCs and 15 VCs, age:32–79y; mean: 53y; f:22; m:8) The clinical diagnosis was based on the criteria of Brandt (1996). Additional inclusion criteria were presentation of symptoms on the day of testing, symptoms present for at least 3 months, no pathology on neuro-ophthalmologic examination and no signs of polyneuropathy. Patients had been medication free for more than 1 month, and they had no history of alcohol abuse. 3. Patients with the presumptive diagnosis of anterior lobe cerebellar atrophy (nZ101; 10 TCs and 10 VCs, age: 13–66y; mean: 59y; f:1; m:19). The diagnosis was based on cerebellar symptoms detected in the clinical neurological examination and atrophy in the vermal and anterior lobe region of the cerebellum on MRI. 4. Patients with the presumptive diagnosis of primary orthostatic tremor (nZ59; 10 TCs and 10 VCs, age: 55–73y; mean: 65y; f:11; m:9). The clinical diagnosis was based on criteria according to Bronstein and Guerraz (1999). 5. Patients with the presumptive diagnosis of unilateral vestibular neuritis (nZ89; 10 TCs and 10 VCs, age: 22–75 y; mean: 56 y; f:5; m:15). This diagnosis was based on the criteria according to Strupp and Brandt (1999).
2.2. Posturography Fore-aft (y), lateral (x) body sway (center of pressure (COP) in mm), and body weight (z, force in Nm) were measured during upright stance (feet next to each other, splayed at an angle of 308, arms always in hanging position; during tandem stance both feet are in a line, one directly behind the other) on a stabilometer platform (Kistler, type 9261A). Sway was recorded in segments of 30-s duration for off-line analysis (sampling frequency 40 Hz). After offset PnK1 elimination, pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi the sway path (SP; e.g. SpxZ iZ1 ðxiC1 Kxi Þ2 for details see Hufschmidtpetffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi al.,P1980)ffi and root mean square (RMS; e.g. RMSxZ 1=n niZ1 x2i ; for details see Brandt et al., 1981) were analyzed for x and y directions. After the Hamming window was applied, Discrete3 Fourier analysis (fft, customized software MATLABw, The MathWorks, Inc., Natick, MA) was used to quantify the distribution of frequencies in the frequency spectrum of body sway. Furthermore, the sum activity of body sway in different frequency ranges was determined as the integral (S(xi)) of the frequency spectrum in the specified range. Frequency ranges of interest were predefined as follows: low frequency range (FFTl: 0.1–2.4 Hz, dominant in normal subjects); middle frequency range
1694
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
(FFTm: 2.43–3.5 Hz, to allow the identification of patients with anterior lobe cerebellar atrophy (3 Hz sway) and PPV patients; high frequency range (FFTh: 3.53–8.0 Hz, to allow the identification of patients with PPV); and very high frequency range (FFTvh: 11–19 Hz, to allow the identification of patients with orthostatic tremor). All patients in the database were assessed according to the same recording protocol. Conditions 1–4 entailed standing on a firm foot support with (1) (2) (3) (4)
eyes eyes eyes eyes
open, closed, open, head extended backward, closed, head extended backward.
Conditions (5), (6), (7), and (8) were the same as conditions 1–4, but a slab of foam rubber (height 10 cm, specific weight 40 g/dm3, stiffness 3.7 kPa) was placed under the rigid foot support. In condition (9) patients had their eyes open and stood in a tandem stance on foam rubber (for details see Querner et al., 2000). Condition (10) was the same as (9) but the eyes were closed. Some patients were not able to perform all 10 conditions because their instability increased and they tended to fall when the condition became more difficult (see Section 3). 2.3. The design of sway patterns as a confidence plot To allow visual classification of the sway patterns, the calculated values of SP, RMS, and FFT were transformed into confidence plots. From each patient’s trial we selected the values (nZ18): SPx, SPy, SPz, RMSx, RMSy, RMSz, FFTlx, FFTly, FFTlz, FFTmx, FFTmy, FFTmz, FFThx, FFThy, FFThz, FFTvhx, FFTvhy, FFTvhz (m, middle; h, high; vh, very high). To simplify the presentation, the patient’s differences from normal subject means were calculated and presented as multiples of the standard deviation (SD) above the mean. This was represented in the confidence plots by a color scale (Fig. 2). For example, pink indicated patient values outside the range of the normal means more than C3SD, red indicated patient values outside the range of the normal meanC10SD and dark blue, differences by less than normal meanC1SD. These confidence plots and the patients histories were used to identify training cases (TCs) and validation cases (VCs). The TCs were used to train a standard back-propagation neural network; VCs were used to test the reliability of the ANNW. 2.4. Standard feed-forward back-propagation neural network In order to train a feed-forward neural network classifier, the back-propagation algorithm was applied according to
Duda et al. (2001). Our ANNW application was implemented using customized software based on the MATLABw (The MathWorks, Inc., Natick, MA) ‘neural network toolbox’. From the calculated parameters of each single condition (such as SP, RMS, FFT) 16 values were selected. The values of 10 conditions for each patient were arranged in a column vector for use as input to the neural network. The components of this vector (input neurons) were consecutively rescaled so as to define all values in the database to lie within the interval of K1 and 1. In detail: for each category of values (e.g. RMSx) the highest and lowest values over all patients were set to 1 and K1, respectively. All other values of the defined category were linearly normalized in relation to these boundary values. As recommended in the recent methodological review (Dreiseitl and Ohno-Machado, 2002), a 3-layer network was chosen as the standard back-propagation neural network (see Fig. 1). All input neurons (n1; m1Z160;1 (16 values! 10 conditions)) were connected to a hidden layer vector (n2; m2Z15;1), whose 15 neurons were all connected to the output vector (n3; m3Z5;1). As result, the output vector n3 gives 5 (patients and normals) different diagnoses between which the network has to differentiate. The multiplication factors of the connections between the layers are the so-called weightings. These weightings are adjusted in the training procedure so as to reach the desired output vector within an acceptable error. The output vector indicates the probability of the suspected diagnosis, presuming the sum of the outputs is 1.
Fig. 1. Design of the standard feed-forward back-propagation neural network after adjustment with the software tool ‘MATLAB neural network toolbox’. The input vector is composed of 160 parameters (16 values!10 posturographic conditions); for the sake of clarity only 4 input variables are depicted here schematically. The hidden layer consists of 15 neurons. In agreement with the number of diagnoses to be differentiated, the output vector has 5 neurons. In the training procedure all 10 conditions may be used to train the net, thus increasing the complexity of the layer connections (not depicted here).
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
In the back-propagation calculation, the weights begin at random values with a mean of zero and then ‘move’ the calculated output vector consecutively step by step in the direction of the desired output vector. To prevent the training procedure from oscillating, optimal adjustments were calculated in one-iteration steps for each patient but were not directly applied to the network. The sum of these adjustments was finally applied to the network, and the next iteration considering again all patients was performed (this regime is also called ‘batch training’). The procedure is usually repeated until the error function between desired and calculated output falls below a predefined threshold (a sum squared error less than 10K6 in this work), thus avoiding an over-fitting (Dreiseitl and Ohno-Machado, 2002), or until a certain number of steps is exceeded (in our case 3000 iterations). Due to our careful selection of training cases and input variables, one iteration converged after approximately 200 steps. To improve the performance of the final classifier, an ensemble of 100 classifiers trained with different initial random weight assignments was used. The output vectors of these networks were combined by taking their mean. After the network had been trained, weight assignments were kept fixed, and the performance of the network was evaluated using our set of validation cases. In this study we focused on the number of correct diagnoses as an indicator of the reliability of the ANNW. Finally, we demonstrated the accuracy of the ANNW by evaluating 5 new cases, which were not included in either group of TCs or VCs. To evaluate the ability of the ANNW the sensitivity and the specificity of the ANNW results were finally calculated. The sensitivity u gives the proportion of the patients with the disease who are correctly identified by the ANNW uZ a/(aCc). aCc is the sum of patients who have the disease, a with positive test results (true positives) and c with negative test results (false negatives). The specificity w gives the proportion of the patients without the disease who are correctly identified by the test. wZd/(bCd). bCd is the sum of patients who do not have the disease, d with negative test results (true negatives) and b with positive test results (false positives).
1695
3.1. Training procedure As intended, the training procedure assigned the correct diagnoses in all TCs with a probability of 1. Fig. 2 shows the color-coded confidence plots as differences between the patients’ (TCs) means and the normal subjects means in multiples of the normal subjects’ standard deviations (SD) of the calculated SP, RMS, and FFT activities. The confidence plot indicates the pathology in patients with PPV by higher SP values and the higher sway activity in the 2.5–8 Hz frequency band (Fig. 2 II). Cerebellar atrophy had dramatically enhanced sway path values and exhibited the highest activity in the 2.5–3 Hz frequency band (Fig. 2 III). Patients with orthostatic tremor showed peak sway activity in the 11–19 Hz frequency band (Fig. 2 IV). Pathologically increased sway values were observed in patients with vestibular neuritis, mainly when their eyes were closed (Fig. 2 V). 3.2. Validation procedure The validation (test) procedure was performed using 60 VCs as characterized in detail in the Section 2. The classification of posturographic sway patterns with the aid of ANNW not only allowed us to distinguish between normals and patients but also to differentiate the 4 postural disorders with an overall sensitivity of 0.93. The overall specificity turned out to also be 0.93. Individual group results were as follows:
3. Results
1. normal subjects: sensitivityZ1, specificityZ1. 2. PPV: sensitivityZ0.87, specificityZ0.96 (one PPV patient was classified as having cerebellar atrophy, another PPV patient, diagnosed as having vestibular neuritis, and two patients with vestibular neuritis were classified as having PPV). 3. Cerebellar atrophy: sensitivityZ1, specificityZ0.98. (1 PPV patient was classified as having a diagnosis of cerebellar atrophy). 4. Orthostatic tremor: sensitivityZ1, specificityZ1. 5. Vestibular neuritis: sensitivityZ0.8, specificityZ0.98 (two patients with vestibular neuritis were classified as having PPV; one PPV patient was diagnosed to have vestibular neuritis).
In order to keep all TCs or VCs in the study, the parameters of only the first 7 measurements were fed into the ANNW. This was necessary, because the 10 conditions made increasing balance demands that required the experimenter to intervene in some cases of impending risk of falls in conditions 8 and 10, especially in patients with cerebellar atrophy, vestibular neuritis, or orthostatic tremor. In some of these TCs and VCs, the experimenter had to actively hold the patients to prevent them from falling. Such conditions were excluded from further analysis.
To exemplify the reliability of ANNW in the clinical routine, the data of 5 individuals are shown in Fig. 3. For the normal subject, the patient with orthostatic tremor, and the patient with vestibular neuritis the accordant output variable was 1, whereas the patients with PPV and cerebellar atrophy were identified with the output variables 0.82 and 0.73, respectively. For the patient with cerebellar atrophy the output variable for the differential diagnosis with vestibular neuritis was calculated to be 0.57.
1696
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
Fig. 2. Color-coded confidence plots of the selected training cases. The abscissa presents differences between the patients’ (TCs) means and the normal subjects’ means in multiples of the normal subjects’ standard deviations (SD) of the calculated SP, RMS, and FFT activities. A value of 57, for example, indicates that the patients’ mean is higher than the normal subjects’ mean C57!SD of the normal subjects’ mean. The color code is as follows. The range of 1!SD!2 is still blue, the range of 2!SD! 3 is violet, the range of 4!SD! 9 is pink; and SDO10 is red. The ordinate represents the conditions of posturographic measurements from 1 to 10.
4. Discussion This study demonstrates that artificial neural network techniques can be used to differentiate postural sway patterns typical of several distinct clinical balance disorders with sufficiently high sensitivity and specificity. To the best of our knowledge this is the first report on an attempt to use ANNW for the analysis of posturographic data. As to the methods, up to now pattern classification has been typically used to compare specific patterns established by ‘learning processes’ with previously unseen patterns. For medical purposes, for example, this technique was successfully applied for the histological identification of tumors in urology (Loch et al., 1999). The possibilities of applying ANNW in the area of clinical biomechanics,
especially in gait analysis, were recently shown in the review of Scho¨llhorn (2004). The major precondition for a reliable application of ANNW is a large database of patients, which allows the identification of a sufficient number of training and validation cases. Moreover, the data must have been obtained under standardized and identical conditions. Once trained, the ANNW delivers an output vector (probability vector) near to 1 (100%) for the diagnosis in the single TCs. This goal is achieved using the backpropagation algorithm, in which errors between desired output and calculated output vector are passed back to the neuron layers to adjust the weightings in order to minimize the errors (Jeffery and Reid, 1997). Then the reliability of the ANNW must be verified by VCs before applying it as a
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
1697
Fig. 3. Application of the artificial neural network to 5 different patients (4 disorders and normal state) after training and validation processes. Left column: original recordings of the body sway in the fore-aft direction (standing on foam, eyes closed). Middle column: color-coded confidence plots. As in Fig. 2 the values are differences between the single new patient’s value and the normal subject’s means in multiples of the normal subject’s standard deviations (SD) of the calculated SP, RMS, and FFT activities. The color-code is the same as in Fig. 2; right column: output values of the ANNW of the particular diagnoses of 5 individuals (N, normal subject; PPV, phobic postural vertigo; CA, cerebellar atrophy; OT, orthostatic tremor; VN, vestibular neuritis).
diagnostic tool for new patients with various balance disorders. The precision of the ANNW can be enhanced by using multiple tests that make different demands on different sensory systems. For example, standing on foam rubber with the eyes closed excludes the influence of vision and decreases the amount of information from lower leg somatosensory afferents but increases the particular sensorial weight of vestibular information. If the results
of the validation procedure are not acceptable (e.g. sensitivity and/or specificity !0.7), the parameters of the network have to be redesigned (Ohno-Machado and Rowland, 1999). To verify its accuracy and reliability, the ANNW in our study was trained a second time by interchanging the groups of TCs and VCs. This did not considerably affect the overall sensitivity (0.93–0.91) or the overall specificity (0.93–0.92).
1698
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
In another analysis the accuracy and reliability of the ANNW were tested after TCs and VCs had been randomly selected from the 5 diagnosis groups (a total of 120 cases) by a co-worker previously not involved in the study. The ANNW was trained with the random TCs and validated with the random VCs. The sensitivity (0.91) and specificity (0.92) were in the same range as calculated before. In general, we do not believe that it is necessary for each laboratory to implement its own software development for ANNW applications. The mathematical operation performed in a common software tool may be adopted by another laboratory as long as the same hardware, method, and parameter extractions are used. Once the ANNW has been designed and thoroughly tested, it can be thought of as a black box, which each examiner can apply to predict a specific diagnosis even before the clinical examination. As soon as the specific sway patterns for more balance disorders are available, the renewed training of ANNW will lead to a broader diagnostic spectrum. We wish to thank Mrs Judy Benson for copyediting the manuscript. P.V. was supported by an ENS Fellowship for 2005. This study was supported by the Bernstein Center for Computational Neuroscience, Munich.
References Brandt T. Phobic postural vertigo. Neurology 1996;46:1515–9. Brandt T, Krafczyk S, Malsbenden I. Postural imbalance with head extension: improvement by training as a model for ataxia therapy. Ann NY Acad Sci 1981;374:636–49. Bronstein AM, Guerraz M. Visual-vestibular control of posture and gait: physiological mechanisms and disorders. Curr Opin Neurol 1999;12: 5–11.
Dichgans J, Mauritz KH, Allum JH, Brandt T. Postural sway in normals and atactic patients: analysis of the stabilising and destabilizing effects of vision. Agressologie 1976;17:15–24. Diener HC, Dichgans J, Bacher M, Gompf B. Quantification of postural sway in normals and patients with cerebellar diseases. Electroencephalogr Clin Neurophysiol 1984;57:134–42. Dreiseitl S, Ohno-Machado L. Logistic regression and artificial neural network classification models: a methodology review. J Biomed Inform 2002;35:352–9. Duda RO, Hart PE, Stork DG. Pattern classification. New York: Wiley– Interscience; 2001. Hahn ME, Chou LS. A model for detecting balance impairment and estimating falls risk in the elderly. Ann Biomed Eng 2005;33:811–20. Hufschmidt A, Dichgans J, Mauritz KH, Hufschmidt M. Some methods and parameters of body sway quantification and their neurological applications. Arch Psychiatr Nervenkr 1980;228:135–50. Jeffery KJ, Reid IC. Modifiable neuronal connections: an overview for psychiatrists. Am J Psychiatry 1997;154:156–64. Krafczyk S, Schlamp V, Dieterich M, Haberhauer P, Brandt T. Increased body sway at 3.5-8 Hz in patients with phobic postural vertigo. Neurosci Lett 1999;259:149–52. Loch T, Leuschner I, Genberg C, Weichert-Jacobsen K, Kuppers F, Yfantis E, Evans M, Tsarev V, Stockle M. Artificial neural network analysis (ANNA) of prostatic transrectal ultrasound. Prostate 1999;39: 198–204. Ohno-Machado L, Rowland T. Neural network applications in physical medicine and rehabilitation. Am J Phys Med Rehabil 1999;78:392–8. Querner V, Krafczyk S, Dieterich M, Brandt T. Patients with somatoform phobic postural vertigo: the more difficult the balance task, the better the balance performance. Neurosci Lett 2000;285:21–4. Scho¨llhorn WI. Applications of artificial neural nets in clinical biomechanics. Clin Biomech (Bristol., Avon.) 2004;19:876–98. Strupp M, Arbusow V, Maag KP, Gall C, Brandt T. Vestibular exercises improve central vestibulospinal compensation after vestibular neuritis. Neurology 1998;51:838–44. Strupp M, Brandt T. Vestibular neuritis. Adv Otorhinolaryngol 1999;55: 111–36. Yarrow K, Brown P, Gresty MA, Bronstein AM. Force platform recordings in the diagnosis of primary orthostatic tremor. Gait Posture 2001;13: 27–34.
Artificial neural network: A new diagnostic posturographic tool for disorders of stance Siegbert Krafczyk a,*, Simon Tietze a, Walter Swoboda a, Peter Valkovicˇ a,b, Thomas Brandt a a b
Department of Neurology, University of Munich, Marchioninistraße 15, D-81377 Munich, Germany 2nd Department of Neurology, School of Medicine, Comenius University, Bratislava, Slovak Republic Accepted 27 April 2006
Abstract Objective: To determine the accuracy of diagnoses made with artificial neural network techniques (ANNW) that identify postural sway patterns typical for balance disorders. Methods: Body sway was measured by means of posturography during 10 test conditions of increasing difficulty. From a database of 676 subjects 60 training cases (TCs) and 60 validation cases (VCs) were selected in which the following diagnoses had been established clinically: normal subject (NS), postural phobic vertigo (PPV), anterior lobe cerebellar atrophy (CA), primary orthostatic tremor (OT), and acute unilateral vestibular neuritis (VN). A standard 3-layer feed-forward ANNW, using the backpropagation algorithm, was trained with TCs, validated with VCs, and its accuracy tested on 5 new cases. Results: ANNW differentiated the established diagnoses with an overall sensitivity and specificity of 0.93. Sensitivity and specificity were 1 for NS and OT; for PPV, 0.87 and 0.96; for CA, 1 and 0.98; and for VN, 0.8 and 0.98, respectively. New subjects were identified with ANNW output variables of the true diagnoses between 0.73 and 1. Conclusions: ANNW differentiates postural sway patterns of several distinct clinical balance disorders with high sensitivity and specificity. Once designed and tested ANNW could be considered a black box, which each examiner can apply to predict a specific diagnosis even without a clinical examination. Significance: A promising diagnostic tool for disorders of upright stance in selected neurological disorders. q 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. Keywords: Upright stance; Posturography; Artificial neural network; Neurological diseases; Back-propagation-algorithm
1. Introduction Static posturography is used to analyze the body sway of unsteady patients with various neurological disorders. Until now clinicians have accepted this method as a tool for following-up balance disorders but not for reliably establishing a specific diagnosis. Their main reason has been that there are no specific analysis criteria of pathological sway patterns. So far diagnoses of only 3 * Corresponding author. Address: Department of Neurology, Klinikum Grosshadern, Ludwig-Maximilians-University, Marchioninistrasse 15, D-81377 Munich, Germany. Tel.: C49 89 7095 4803; fax: C49 89 7095 4801. E-mail address: [email protected] (S. Krafczyk).
conditions have been established on the basis of a routine analysis of body sway: (i) the 3 Hz sway in anterior lobe cerebellar atrophy (Dichgans et al., 1976; Diener et al., 1984); (ii) increased sway activity in the higher power spectra frequency band with a typical peak between 12 and 19 Hz in primary orthostatic tremor patients (Bronstein and Guerraz, 1999; Yarrow et al., 2001); and (iii) increased sway activity in the 3.5–8 Hz frequency band in patients with somatoform phobic postural vertigo (PPV) (Krafczyk et al., 1999). It is much more difficult to determine posturographic criteria for vestibular disorders like vestibular neuritis. The body sway of these patients increases in darkness during the first days after disease onset; however, because of the large variation in balance performance it cannot be differentiated from other balance disorders by
1388-2457/$30.00 q 2006 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.clinph.2006.04.022
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
body sway alone (Strupp and Brandt, 1999; Strupp et al., 1998). Basically, however, it is possible to use a single or a combination of multiple parameters to characterize a certain balance disorder. In the following study we improved the discriminatory power of posturography by using multiple parameters obtained from the original data. These included sway path (SP) (for details see Hufschmidt et al., 1980); root mean square (RMS) (Brandt et al., 1981); and the sum activity of body sway in different frequency ranges after Fourier transformation analysis (FFT). Our goal was to determine the typical patterns of pathological sway for 4 disorders of upright stance: anterior lobe cerebellar atrophy, PPV, primary orthostatic tremor, and acute vestibular neuritis. Since, this approach increases the complexity of the data comparison to such an extent that it is impractical for routine clinical use, an artificial neural network (ANNW), as described by Duda et al. (2001), was applied for computational analysis of the sway patterns in the patients and in normal controls. The advantage of ANNW is that all posturographic parameters, as well as many pre-selected sway patterns, which have been identified by experienced examiners, can be included. The effectiveness of an ANNW model was newly shown by classifying the risk of falls in the elderly on the basis of an analysis of balance control during gait (Hahn and Chou, 2005). We recorded the postural sway characteristics of 4 neurological and vestibular disorders of upright stance and fed an ANNW with multiple parameters extracted from the raw data. Initially, the ANNW was trained with the sway parameters of 60 training cases (TCs) and was then validated with 60 validation cases (VCs). The VCs included 5 categories (4 balance disorders and normal sway); the diagnoses had been previously established clinically. The goal was to test the validity of the ANNW when used to determine the probable diagnosis of a balance disorder by comparing the data of new patients with the stored experience of the ANNW but without any access to further information.
2. Material and Methods 2.1. Patients A database of 676 adult patients, who had been assessed according to the same protocol and had undergone static posturography between 1998 and 2005, served as the basis for selecting 421 patients in the categories under investigation. The TCs and VCs (normal subjects included) were selected after the identification of congruent sway patterns with the help of confidence plots (see Section 2.3) and the verification of the diagnosis by clinical methods on the basis of the patients’ data sheet after discharge from the clinic and the establishment of the final diagnosis. The following
1693
groups were selected for the pattern classification of posturographic sway: 1. normal subjects (total number in the database, nZ41; 15 TCs and 15 VCs were selected, age:18–54y; mean: 29y; f:11; m:19). 2. Patients with the presumptive diagnosis of PPV (nZ225; 15 TCs and 15 VCs, age:32–79y; mean: 53y; f:22; m:8) The clinical diagnosis was based on the criteria of Brandt (1996). Additional inclusion criteria were presentation of symptoms on the day of testing, symptoms present for at least 3 months, no pathology on neuro-ophthalmologic examination and no signs of polyneuropathy. Patients had been medication free for more than 1 month, and they had no history of alcohol abuse. 3. Patients with the presumptive diagnosis of anterior lobe cerebellar atrophy (nZ101; 10 TCs and 10 VCs, age: 13–66y; mean: 59y; f:1; m:19). The diagnosis was based on cerebellar symptoms detected in the clinical neurological examination and atrophy in the vermal and anterior lobe region of the cerebellum on MRI. 4. Patients with the presumptive diagnosis of primary orthostatic tremor (nZ59; 10 TCs and 10 VCs, age: 55–73y; mean: 65y; f:11; m:9). The clinical diagnosis was based on criteria according to Bronstein and Guerraz (1999). 5. Patients with the presumptive diagnosis of unilateral vestibular neuritis (nZ89; 10 TCs and 10 VCs, age: 22–75 y; mean: 56 y; f:5; m:15). This diagnosis was based on the criteria according to Strupp and Brandt (1999).
2.2. Posturography Fore-aft (y), lateral (x) body sway (center of pressure (COP) in mm), and body weight (z, force in Nm) were measured during upright stance (feet next to each other, splayed at an angle of 308, arms always in hanging position; during tandem stance both feet are in a line, one directly behind the other) on a stabilometer platform (Kistler, type 9261A). Sway was recorded in segments of 30-s duration for off-line analysis (sampling frequency 40 Hz). After offset PnK1 elimination, pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi the sway path (SP; e.g. SpxZ iZ1 ðxiC1 Kxi Þ2 for details see Hufschmidtpetffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi al.,P1980)ffi and root mean square (RMS; e.g. RMSxZ 1=n niZ1 x2i ; for details see Brandt et al., 1981) were analyzed for x and y directions. After the Hamming window was applied, Discrete3 Fourier analysis (fft, customized software MATLABw, The MathWorks, Inc., Natick, MA) was used to quantify the distribution of frequencies in the frequency spectrum of body sway. Furthermore, the sum activity of body sway in different frequency ranges was determined as the integral (S(xi)) of the frequency spectrum in the specified range. Frequency ranges of interest were predefined as follows: low frequency range (FFTl: 0.1–2.4 Hz, dominant in normal subjects); middle frequency range
1694
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
(FFTm: 2.43–3.5 Hz, to allow the identification of patients with anterior lobe cerebellar atrophy (3 Hz sway) and PPV patients; high frequency range (FFTh: 3.53–8.0 Hz, to allow the identification of patients with PPV); and very high frequency range (FFTvh: 11–19 Hz, to allow the identification of patients with orthostatic tremor). All patients in the database were assessed according to the same recording protocol. Conditions 1–4 entailed standing on a firm foot support with (1) (2) (3) (4)
eyes eyes eyes eyes
open, closed, open, head extended backward, closed, head extended backward.
Conditions (5), (6), (7), and (8) were the same as conditions 1–4, but a slab of foam rubber (height 10 cm, specific weight 40 g/dm3, stiffness 3.7 kPa) was placed under the rigid foot support. In condition (9) patients had their eyes open and stood in a tandem stance on foam rubber (for details see Querner et al., 2000). Condition (10) was the same as (9) but the eyes were closed. Some patients were not able to perform all 10 conditions because their instability increased and they tended to fall when the condition became more difficult (see Section 3). 2.3. The design of sway patterns as a confidence plot To allow visual classification of the sway patterns, the calculated values of SP, RMS, and FFT were transformed into confidence plots. From each patient’s trial we selected the values (nZ18): SPx, SPy, SPz, RMSx, RMSy, RMSz, FFTlx, FFTly, FFTlz, FFTmx, FFTmy, FFTmz, FFThx, FFThy, FFThz, FFTvhx, FFTvhy, FFTvhz (m, middle; h, high; vh, very high). To simplify the presentation, the patient’s differences from normal subject means were calculated and presented as multiples of the standard deviation (SD) above the mean. This was represented in the confidence plots by a color scale (Fig. 2). For example, pink indicated patient values outside the range of the normal means more than C3SD, red indicated patient values outside the range of the normal meanC10SD and dark blue, differences by less than normal meanC1SD. These confidence plots and the patients histories were used to identify training cases (TCs) and validation cases (VCs). The TCs were used to train a standard back-propagation neural network; VCs were used to test the reliability of the ANNW. 2.4. Standard feed-forward back-propagation neural network In order to train a feed-forward neural network classifier, the back-propagation algorithm was applied according to
Duda et al. (2001). Our ANNW application was implemented using customized software based on the MATLABw (The MathWorks, Inc., Natick, MA) ‘neural network toolbox’. From the calculated parameters of each single condition (such as SP, RMS, FFT) 16 values were selected. The values of 10 conditions for each patient were arranged in a column vector for use as input to the neural network. The components of this vector (input neurons) were consecutively rescaled so as to define all values in the database to lie within the interval of K1 and 1. In detail: for each category of values (e.g. RMSx) the highest and lowest values over all patients were set to 1 and K1, respectively. All other values of the defined category were linearly normalized in relation to these boundary values. As recommended in the recent methodological review (Dreiseitl and Ohno-Machado, 2002), a 3-layer network was chosen as the standard back-propagation neural network (see Fig. 1). All input neurons (n1; m1Z160;1 (16 values! 10 conditions)) were connected to a hidden layer vector (n2; m2Z15;1), whose 15 neurons were all connected to the output vector (n3; m3Z5;1). As result, the output vector n3 gives 5 (patients and normals) different diagnoses between which the network has to differentiate. The multiplication factors of the connections between the layers are the so-called weightings. These weightings are adjusted in the training procedure so as to reach the desired output vector within an acceptable error. The output vector indicates the probability of the suspected diagnosis, presuming the sum of the outputs is 1.
Fig. 1. Design of the standard feed-forward back-propagation neural network after adjustment with the software tool ‘MATLAB neural network toolbox’. The input vector is composed of 160 parameters (16 values!10 posturographic conditions); for the sake of clarity only 4 input variables are depicted here schematically. The hidden layer consists of 15 neurons. In agreement with the number of diagnoses to be differentiated, the output vector has 5 neurons. In the training procedure all 10 conditions may be used to train the net, thus increasing the complexity of the layer connections (not depicted here).
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
In the back-propagation calculation, the weights begin at random values with a mean of zero and then ‘move’ the calculated output vector consecutively step by step in the direction of the desired output vector. To prevent the training procedure from oscillating, optimal adjustments were calculated in one-iteration steps for each patient but were not directly applied to the network. The sum of these adjustments was finally applied to the network, and the next iteration considering again all patients was performed (this regime is also called ‘batch training’). The procedure is usually repeated until the error function between desired and calculated output falls below a predefined threshold (a sum squared error less than 10K6 in this work), thus avoiding an over-fitting (Dreiseitl and Ohno-Machado, 2002), or until a certain number of steps is exceeded (in our case 3000 iterations). Due to our careful selection of training cases and input variables, one iteration converged after approximately 200 steps. To improve the performance of the final classifier, an ensemble of 100 classifiers trained with different initial random weight assignments was used. The output vectors of these networks were combined by taking their mean. After the network had been trained, weight assignments were kept fixed, and the performance of the network was evaluated using our set of validation cases. In this study we focused on the number of correct diagnoses as an indicator of the reliability of the ANNW. Finally, we demonstrated the accuracy of the ANNW by evaluating 5 new cases, which were not included in either group of TCs or VCs. To evaluate the ability of the ANNW the sensitivity and the specificity of the ANNW results were finally calculated. The sensitivity u gives the proportion of the patients with the disease who are correctly identified by the ANNW uZ a/(aCc). aCc is the sum of patients who have the disease, a with positive test results (true positives) and c with negative test results (false negatives). The specificity w gives the proportion of the patients without the disease who are correctly identified by the test. wZd/(bCd). bCd is the sum of patients who do not have the disease, d with negative test results (true negatives) and b with positive test results (false positives).
1695
3.1. Training procedure As intended, the training procedure assigned the correct diagnoses in all TCs with a probability of 1. Fig. 2 shows the color-coded confidence plots as differences between the patients’ (TCs) means and the normal subjects means in multiples of the normal subjects’ standard deviations (SD) of the calculated SP, RMS, and FFT activities. The confidence plot indicates the pathology in patients with PPV by higher SP values and the higher sway activity in the 2.5–8 Hz frequency band (Fig. 2 II). Cerebellar atrophy had dramatically enhanced sway path values and exhibited the highest activity in the 2.5–3 Hz frequency band (Fig. 2 III). Patients with orthostatic tremor showed peak sway activity in the 11–19 Hz frequency band (Fig. 2 IV). Pathologically increased sway values were observed in patients with vestibular neuritis, mainly when their eyes were closed (Fig. 2 V). 3.2. Validation procedure The validation (test) procedure was performed using 60 VCs as characterized in detail in the Section 2. The classification of posturographic sway patterns with the aid of ANNW not only allowed us to distinguish between normals and patients but also to differentiate the 4 postural disorders with an overall sensitivity of 0.93. The overall specificity turned out to also be 0.93. Individual group results were as follows:
3. Results
1. normal subjects: sensitivityZ1, specificityZ1. 2. PPV: sensitivityZ0.87, specificityZ0.96 (one PPV patient was classified as having cerebellar atrophy, another PPV patient, diagnosed as having vestibular neuritis, and two patients with vestibular neuritis were classified as having PPV). 3. Cerebellar atrophy: sensitivityZ1, specificityZ0.98. (1 PPV patient was classified as having a diagnosis of cerebellar atrophy). 4. Orthostatic tremor: sensitivityZ1, specificityZ1. 5. Vestibular neuritis: sensitivityZ0.8, specificityZ0.98 (two patients with vestibular neuritis were classified as having PPV; one PPV patient was diagnosed to have vestibular neuritis).
In order to keep all TCs or VCs in the study, the parameters of only the first 7 measurements were fed into the ANNW. This was necessary, because the 10 conditions made increasing balance demands that required the experimenter to intervene in some cases of impending risk of falls in conditions 8 and 10, especially in patients with cerebellar atrophy, vestibular neuritis, or orthostatic tremor. In some of these TCs and VCs, the experimenter had to actively hold the patients to prevent them from falling. Such conditions were excluded from further analysis.
To exemplify the reliability of ANNW in the clinical routine, the data of 5 individuals are shown in Fig. 3. For the normal subject, the patient with orthostatic tremor, and the patient with vestibular neuritis the accordant output variable was 1, whereas the patients with PPV and cerebellar atrophy were identified with the output variables 0.82 and 0.73, respectively. For the patient with cerebellar atrophy the output variable for the differential diagnosis with vestibular neuritis was calculated to be 0.57.
1696
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
Fig. 2. Color-coded confidence plots of the selected training cases. The abscissa presents differences between the patients’ (TCs) means and the normal subjects’ means in multiples of the normal subjects’ standard deviations (SD) of the calculated SP, RMS, and FFT activities. A value of 57, for example, indicates that the patients’ mean is higher than the normal subjects’ mean C57!SD of the normal subjects’ mean. The color code is as follows. The range of 1!SD!2 is still blue, the range of 2!SD! 3 is violet, the range of 4!SD! 9 is pink; and SDO10 is red. The ordinate represents the conditions of posturographic measurements from 1 to 10.
4. Discussion This study demonstrates that artificial neural network techniques can be used to differentiate postural sway patterns typical of several distinct clinical balance disorders with sufficiently high sensitivity and specificity. To the best of our knowledge this is the first report on an attempt to use ANNW for the analysis of posturographic data. As to the methods, up to now pattern classification has been typically used to compare specific patterns established by ‘learning processes’ with previously unseen patterns. For medical purposes, for example, this technique was successfully applied for the histological identification of tumors in urology (Loch et al., 1999). The possibilities of applying ANNW in the area of clinical biomechanics,
especially in gait analysis, were recently shown in the review of Scho¨llhorn (2004). The major precondition for a reliable application of ANNW is a large database of patients, which allows the identification of a sufficient number of training and validation cases. Moreover, the data must have been obtained under standardized and identical conditions. Once trained, the ANNW delivers an output vector (probability vector) near to 1 (100%) for the diagnosis in the single TCs. This goal is achieved using the backpropagation algorithm, in which errors between desired output and calculated output vector are passed back to the neuron layers to adjust the weightings in order to minimize the errors (Jeffery and Reid, 1997). Then the reliability of the ANNW must be verified by VCs before applying it as a
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
1697
Fig. 3. Application of the artificial neural network to 5 different patients (4 disorders and normal state) after training and validation processes. Left column: original recordings of the body sway in the fore-aft direction (standing on foam, eyes closed). Middle column: color-coded confidence plots. As in Fig. 2 the values are differences between the single new patient’s value and the normal subject’s means in multiples of the normal subject’s standard deviations (SD) of the calculated SP, RMS, and FFT activities. The color-code is the same as in Fig. 2; right column: output values of the ANNW of the particular diagnoses of 5 individuals (N, normal subject; PPV, phobic postural vertigo; CA, cerebellar atrophy; OT, orthostatic tremor; VN, vestibular neuritis).
diagnostic tool for new patients with various balance disorders. The precision of the ANNW can be enhanced by using multiple tests that make different demands on different sensory systems. For example, standing on foam rubber with the eyes closed excludes the influence of vision and decreases the amount of information from lower leg somatosensory afferents but increases the particular sensorial weight of vestibular information. If the results
of the validation procedure are not acceptable (e.g. sensitivity and/or specificity !0.7), the parameters of the network have to be redesigned (Ohno-Machado and Rowland, 1999). To verify its accuracy and reliability, the ANNW in our study was trained a second time by interchanging the groups of TCs and VCs. This did not considerably affect the overall sensitivity (0.93–0.91) or the overall specificity (0.93–0.92).
1698
S. Krafczyk et al. / Clinical Neurophysiology 117 (2006) 1692–1698
In another analysis the accuracy and reliability of the ANNW were tested after TCs and VCs had been randomly selected from the 5 diagnosis groups (a total of 120 cases) by a co-worker previously not involved in the study. The ANNW was trained with the random TCs and validated with the random VCs. The sensitivity (0.91) and specificity (0.92) were in the same range as calculated before. In general, we do not believe that it is necessary for each laboratory to implement its own software development for ANNW applications. The mathematical operation performed in a common software tool may be adopted by another laboratory as long as the same hardware, method, and parameter extractions are used. Once the ANNW has been designed and thoroughly tested, it can be thought of as a black box, which each examiner can apply to predict a specific diagnosis even before the clinical examination. As soon as the specific sway patterns for more balance disorders are available, the renewed training of ANNW will lead to a broader diagnostic spectrum. We wish to thank Mrs Judy Benson for copyediting the manuscript. P.V. was supported by an ENS Fellowship for 2005. This study was supported by the Bernstein Center for Computational Neuroscience, Munich.
References Brandt T. Phobic postural vertigo. Neurology 1996;46:1515–9. Brandt T, Krafczyk S, Malsbenden I. Postural imbalance with head extension: improvement by training as a model for ataxia therapy. Ann NY Acad Sci 1981;374:636–49. Bronstein AM, Guerraz M. Visual-vestibular control of posture and gait: physiological mechanisms and disorders. Curr Opin Neurol 1999;12: 5–11.
Dichgans J, Mauritz KH, Allum JH, Brandt T. Postural sway in normals and atactic patients: analysis of the stabilising and destabilizing effects of vision. Agressologie 1976;17:15–24. Diener HC, Dichgans J, Bacher M, Gompf B. Quantification of postural sway in normals and patients with cerebellar diseases. Electroencephalogr Clin Neurophysiol 1984;57:134–42. Dreiseitl S, Ohno-Machado L. Logistic regression and artificial neural network classification models: a methodology review. J Biomed Inform 2002;35:352–9. Duda RO, Hart PE, Stork DG. Pattern classification. New York: Wiley– Interscience; 2001. Hahn ME, Chou LS. A model for detecting balance impairment and estimating falls risk in the elderly. Ann Biomed Eng 2005;33:811–20. Hufschmidt A, Dichgans J, Mauritz KH, Hufschmidt M. Some methods and parameters of body sway quantification and their neurological applications. Arch Psychiatr Nervenkr 1980;228:135–50. Jeffery KJ, Reid IC. Modifiable neuronal connections: an overview for psychiatrists. Am J Psychiatry 1997;154:156–64. Krafczyk S, Schlamp V, Dieterich M, Haberhauer P, Brandt T. Increased body sway at 3.5-8 Hz in patients with phobic postural vertigo. Neurosci Lett 1999;259:149–52. Loch T, Leuschner I, Genberg C, Weichert-Jacobsen K, Kuppers F, Yfantis E, Evans M, Tsarev V, Stockle M. Artificial neural network analysis (ANNA) of prostatic transrectal ultrasound. Prostate 1999;39: 198–204. Ohno-Machado L, Rowland T. Neural network applications in physical medicine and rehabilitation. Am J Phys Med Rehabil 1999;78:392–8. Querner V, Krafczyk S, Dieterich M, Brandt T. Patients with somatoform phobic postural vertigo: the more difficult the balance task, the better the balance performance. Neurosci Lett 2000;285:21–4. Scho¨llhorn WI. Applications of artificial neural nets in clinical biomechanics. Clin Biomech (Bristol., Avon.) 2004;19:876–98. Strupp M, Arbusow V, Maag KP, Gall C, Brandt T. Vestibular exercises improve central vestibulospinal compensation after vestibular neuritis. Neurology 1998;51:838–44. Strupp M, Brandt T. Vestibular neuritis. Adv Otorhinolaryngol 1999;55: 111–36. Yarrow K, Brown P, Gresty MA, Bronstein AM. Force platform recordings in the diagnosis of primary orthostatic tremor. Gait Posture 2001;13: 27–34.