This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier’s archiving and manuscript policies are encouraged to visit: http://www.elsevier.com/copyright Author's personal copy Journal of Neuroscience Methods 171 (2008) 264–270 Contents lists available at ScienceDirect Journal of Neuroscience Methods journal homepage: www.elsevier.com/locate/jneumeth Spiral analysis—Improved clinical utility with center detection Hongzhi Wang a,b , Qiping Yu a , Mónica M. Kurtis a , Alicia G. Floyd a , Whitney A. Smith a , Seth L. Pullman a,∗ a Clinical Motor Physiology Laboratory, Department of Neurology, Columbia University Medical Center, The Neurological Institute, 710 West 168th Street, New York, NY 10032, United States b Computer Vision Laboratory, Department of Computer Science, Stevens Institute of Technology, Hoboken, NJ, United States a r t i c l e i n f o Article history: Received 26 September 2007 Received in revised form 21 March 2008 Accepted 21 March 2008 Keywords: Spiral analysis Movement disorders Parkinson’s disease Essential tremor Dystonia Motor performance Curve fitting a b s t r a c t Spiral analysis is a computerized method that measures human motor performance from handwritten Archimedean spirals. It quantifies normal motor activity, and detects early disease as well as dysfunction in patients with movement disorders. The clinical utility of spiral analysis is based on kinematic and dynamic indices derived from the original spiral trace, which must be detected and transformed into mathematical expressions with great precision. Accurately determining the center of the spiral and reducing spurious low frequency noise caused by center selection error is important to the analysis. Handwritten spirals do not all start at the same point, even when marked on paper, and drawing artifacts are not easily filtered without distortion of the spiral data and corruption of the performance indices. In this report, we describe a method for detecting the optimal spiral center and reducing the unwanted drawing artifacts. To demonstrate overall improvement to spiral analysis, we study the impact of the optimal spiral center detection in different frequency domains separately and find that it notably improves the clinical spiral measurement accuracy in low frequency domains. © 2008 Elsevier B.V. All rights reserved. 1. Introduction Handwriting or drawing can be used to study physiologic issues in normal controls (NCs) and in patients with movement disorders such as Parkinson’s disease (PD), essential tremor (ET), and dystonia (DY) (Elble et al., 1990, 1996; Forsyth and Ponce, 2003; Hogg et al., 2005; Liu et al., 2005; Louis et al., 2006; Siebner et al., 1999). The most commonly measured parameters relate to the kinematics and dynamics of writing and drawing. Compared to handwriting, drawing requires subjects to execute standard figures, e.g. straight lines, circles or spirals so more comparable information, such as drawing shape and consistency, can be evaluated (Elble et al., 1990, 1996; Hogg et al., 2005; Miralles et al., 2006; Pullman, 1998; Wang et al., 2005). Spiral analysis is a clinical test that records Archimedean spirals drawn on a digitizing tablet connected to a computer (Pullman, 1998). It is used to quantify normal motor activity as well as measure dysfunction in patients with movement disorders (Saunders-Pullman et al., 2008). With an inking pen that also functions as a computer mouse, subjects draw spi- ∗ Corresponding author. Tel.: +1 212 305 1331; fax: +1 212 305 0743. E-mail address: sp31@columbia.edu (S.L. Pullman). 0165-0270/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.jneumeth.2008.03.009 rals in 10 cm × 10 cm with the starting point marked. Data are collected in the x, y and z (pressure) axes and provide virtual tri-axial recordings of spiral kinematics and dynamics. This analysis extends spiral drawing, a commonly performed neurologic test, into an objective and accurate measure of motor performance. Spiral analysis extracts specific quantifiable indices from the handwritten spirals, such as linear and angular drawing speed, acceleration, time-related changes in all these measurements, loop width and consistency in the x–y plane, pressure in the z-axis, among many others. Abnormal movements that are superimposed on the spirals such as tremors are also evaluated. Multiple indices can be combined to provide more complex analyses of spiral drawing, including an overall degree of severity. Compared to simple drawing tasks, e.g. a straight line, spiral execution is a multi-jointed task involving the distal and proximal upper limb and thus provides a window on a wide range of motor problems (Hogg et al., 2005; Lacquaniti et al., 1987; Wang et al., 2005). As explained below, an important feature of spiral drawing is that it can be transformed into an expression using polar coordinates with complete fidelity to the original spiral shape. The transformation of spiral data from image to polar coordinates results in a low order polynomial curve representation that simplifies the extraction of the properties of the Archimedean spi- Author's personal copy 265 H. Wang et al. / Journal of Neuroscience Methods 171 (2008) 264–270 Fig. 1. (a) An ideal spiral with the center (0,0) the starting point of the spiral, indicated by the open circle, and another point (−10,−10) indicated by the open square below and left of the true center; (b) the polar transformation of (a) based on the proper center (0,0); (c) the polar transformation of (a) based on the inaccurate center (−10,−10). Note the markedly different transformation curve with spurious low frequency noise. rals. Transformation of the two-dimensional spiral drawing into its polar expression is as follows: x = r sin ! + x0 , r = ! y = r cos ! + y0 ; 2 (x − x0 ) + (y − y0 )2 , arctan ! = y − y0 x − x0 where (x0 , y0 ) is the start of the spiral drawing (x, y) is the spiral image coordinate and (r, !) is its counterpart in polar expression. The Archimedean spiral polar expression has a simple polynomial relationship, i.e. ! = k + arn , where a, n are positive constants and k controls spirals orientation. The polar equations of spirals can be “normalized” to an orientation where k = 0 as in Fig. 1, where a = 5, n = 1. From the above equations, it is evident that the spiral center is critical to their polar transformations. When the center is incorrectly identified, the polar expression becomes irregular (Fig. 1c) with superimposed low frequency oscillations, which may result in less accurately measured spiral indices. Spiral onset can be taken many different ways, for example as the geometric center of the 10 cm × 10 cm outline or as the first nonzero pressure value after the subject starts drawing (Pullman, 1998; Wang et al., 2005). However, many subjects, particularly movement disorder patients, do not execute spirals in a controlled manner at initiation. Rarely do patients start from the center mark making the first non-zero pressure point unreliable as the spiral center. Furthermore, poorly executed spirals and abnormal movements such as tremors, superimposed on the spiral shape, cause artifacts that distort the spiral data, alter the polar coordinates, and result in improperly defined spiral centers. High order polynomials cannot be utilized to define the spiral and its center because any irregular polar expression can be approximated to yield the original, erratically initiated spiral. Without specific restrictions on the polynomial function, any (x, y) coordinate may falsely yield an apparent spiral center. To our knowledge, there is no automatic method to optimize spiral center detection, and reduce low frequency drawing artifacts to improve the clinical utility of computerized spiral analysis. In this report, we describe a regression-based approach to achieve these goals. We then test our technique in a disease classification task to document the improvement in the clinical utility. 2. Experimental methods 2.1. Data collection We obtain 10 freely drawn (no constraints, no attachments and no templates to trace) handwritten spirals from NC subjects (n = 10) and patients with PD, ET and DY (n = 10 for each condition), yielding a total of 400 spirals for the study. All subjects are instructed to draw Archimedean spirals as well as possible inside a 10 cm × 10 cm square on an 8.5 in. × 11 in. white paper mounted on a graphics tablet. The drawing pen is wireless and there are no other attachments to the subjects. The tablet has a resolution of 10 pixels/mm, or an accuracy of 0.1 mm/pixel, an output rate of 200 points/s, and 256 levels of measurable pressure. This provides approximately 70 Hz sampling per channel. The captured spiral data are then normalized to have the same time interval between consecutive points via linear interpolation. 2.2. Optimization through regression analysis The polar coordinates of a spiral may be represented by a curve, R(C) = {ri (C),! i (C)}i = 1:n , given a center C, where i is the ordered index to spiral point and n is the total number of spiral points. With any given center, there are infinite low order ideal Archimedean spirals passing any two consecutive spiral points, but an actual drawn spiral cannot follow the ideal spiral definition exactly due to the irregularities and inconsistencies of drawing different parts of the spiral. We define that the local optimal center of an actual drawn spiral is the one that is the most consistent with all spiral points. Hence, the corresponding optimal spiral should have the least fitting error. We perform the regression analysis in polar domain instead of the image domain to take advantage of its simpler expression. We assign R̂(C) = {ri (C), !ˆ i (C)}i=1:n to be the curve obtained by fitting a low order polynomial to the curve R(C). We use the least square fitting method, i.e. minimize the sum of squared residual error, to obtain the fitting curve. We describe how to test the effects of the order of polynomials in the next section. The optimal center, C̄, gives the minimum fitting error: "n |! (C) − !ˆ i (C)| i=1 i C̄ = arg minC F(C) = arg minC "n−1 |! (C) − !i+1 (C)| i=1 i (1) where F(C) is a normalized curve fitting error for center C. The denominator is for normalization purpose so that F(C) is invariant to the scale of the spiral. For simplicity, we use the least square fitting method, i.e. minimize the sum of squared residual error, to obtain the fitting curve. The curve obtained by least square fitting may not minimize the sum of absolute residual errors, the numerator in Eq. (1), but provides a reasonable approximation. Though one can further refine the curve via gradient descent optimization, we find that the least square curve fitting is sufficient to give good center detection results. The normalized fitting error provides a measure of spiral irregularity. We then use the gradient descent approach to minimize the F(C) function. Adjusting C in the direction of the derivative of function F(C) with respect to the center, dF(C)/dC, allows C to converge at Author's personal copy 266 H. Wang et al. / Journal of Neuroscience Methods 171 (2008) 264–270 an local optimal center corresponding to a local minimum of F(C). For simplicity, instead of the continuous searching space we only search for optimal centers in the discrete image pixel coordinates. The following summarizes our optimization procedure: (1) Initialize the center estimate, C0 = (x0 ,y0 ), to" be the image center " 0 xi /n and y0 = yi /n. of the spiral, i.e. x = i=1:n i=1:n (2) Compute the F(C) for each pixel in the window within a small neighborhood of the estimated center. (3) If the pixel giving the best F(C) in the neighborhood window is different from the current choice, move the center estimate to this pixel, and repeat steps 2 and 3. Otherwise, set the current choice as the local optimal spiral center. The choice of the local searching window size is a compromise between global optimal solution and computational costs. A large local searching window has a better chance of finding the global optimal solution, but at a higher computational cost. In this study, we use a 21 × 21 pixel searching window, or 441 pixels. The optimization requires several seconds to converge using Matlab (The MathWorks, Inc., Natick, MA) with a 2 GHz CPU computer. 2.3. Analysis of computer-generated spirals To test our spiral center detection technique, we apply our algorithm using computer-generated spiral data measured in pixels. The spiral data are synthesized with the polar expression using ! = r˛ , where ˛ = {0.8, 0.9, 1, 1.1, 1.2} and 50 ≤ r ≤ 400 with a constant step of 1 pixel. Thus, each spiral contains 351 sample points. To test the robustness of this method, we deteriorate r and ! by adding random noise evenly distributed over a radius [−5,5], in pixels, and angle [−0.25,0.25], in radians, respectively. Image coordinates are then calculated from the polar coordinates and utilized for test, with the image coordinate of the ground truth spiral center as (0,0). For each ˛, we randomly generate 10 synthesized spirals, resulting in a total of 50 spiral constructions. We defined the center detection error as the Euclidean distance between the detected center and the ground truth. We test the sensitivity of our optimization algorithm by initializing the center estimate to be the image spiral center plus random noise evenly distributed in [−50,50]. We then search for optimal centers using the algorithm outlined above for each spiral using 1st, 2nd, and 3rd order polynomial fitting separately, resulting in a total of 150 trials. 2.4. Multi-scale analysis To demonstrate the utility of our approach, we analyse the impact of the local optimal center at different frequency domains separately. To this end, we first decompose a spiral into different frequency domains using the Gaussian pyramid technique. Then we compare the accuracy of the measured spiral indices, evaluating drawing irregularities, with and without our optimal center at each frequency domain separately. Widely applied in computer vision studies, the Gaussian pyramid technique is a form of image processing that utilizes decreasing pixel resolution representations of the spiral as a form of low-pass filtering. As in center detection, the Gaussian pyramid technique is applied on the polar coordinates. We use the following repeated smoothing and down-sampling procedures (Forsyth and Ponce, 2003; Lindeberg, 1994): (1) Applying the Gaussian kernel method to remove high frequency signals by smoothing the spiral curves. This smoothing was performed to avoid the aliasing effect of the next down-sampling 2 2 procedure. The Gaussian kernel is G(x) = e−x /2" , where " is the deviation. We used " = 2, which as shown in (Yu, 2004) is sufficient to avoid aliasing. (2) Down-sampling the spiral by 2, where the jth element of the down-sampled spiral was the 2jth element of the original spiral. Since the unit in the coarser scale had twice the magnitude unit from the previous iteration, the spiral scale was coarser at half the width and height of the previous drawing, with one-fourth the area. The coarser spiral was visually equivalent to being observed at a distance twice that of the previous scale. Each of these iterations is applied to spirals already processed for optimal center detection, gradually eliminating spuriously low frequency oscillations in the polar domain. As the digitizing tablet samples at approximately 70 Hz per channel, according to the Nyquist theorem (Forsyth and Ponce, 2003), the spirals were recorded without aliasing up to ∼35 Hz. When the sampling frequency is halved in 2nd scale, only signals up to ∼17 Hz are preserved and the differences between 1st and 2nd scale spirals, the radial oscillation speeds described below, contained signals between 17 and 35 Hz. The amplitude of the radial oscillation speeds corresponds to the signal power in this frequency domain. Similarly, the signals preserved in 3rd, 4th and 5th scales are ∼8, ∼4 and ∼2 Hz, respectively. Since tremors and most other clinically abnormal movements are composed of signals with frequencies greater than 2 Hz, the 5th scale spiral is effectively without tremor or abnormal movements. In our analysis, we decompose spirals into five scales. Hence, we need to perform down sampling four times for each spiral. Compared with other global frequency analysis approaches, e.g. Fourier transforms, the Gaussian pyramid technique accurately locates the original signals at different frequency domains in the original spiral (Fig. 3). To measure the difference between the local optimal center and pre-defined center with measured spiral quantities, we compare the separation ability between clinically normal and abnormal spirals drawn by patients with PD, ET and DY. If the local optimal center obtained by our method is better than pre-defined centers, the corresponding measured spiral quantities should be more accurate and give better separation performance. We compute the oscillation speed of tremors and other abnormal movements that superimposed on the processed spirals as spiral quantities to capture the degree of abnormality. Given the radius curve of a spiral {(C)}si=1:n at a scale s in the polar domain, we compute the oscillation speed for each spiral point as follows: (1) Remove the high-frequency abnormality from the spiral curve by a smoothing algorithm using the same Gaussian kernel above. Let the smoothed radius curve be {r̄i (C)}si=1:n . (2) Compute the oscillation amplitude, {ai (C)}si=1:n , as the radial difference between the original and the abnormality free radius curve, i.e. {ai (C) = ri (C) − r̄i (C)}si=1:n . (3) Compute the radial oscillation speeds, {vi (C)}si=1:n , for spiral points as the absolute difference between the radial oscillation amplitudes of consecutive spiral points, i.e. {vi (C) = ai (C) − ai−1 (C)}si=1:n−1 . The points with higher radial oscillation speeds are considered as more likely containing abnormality than those with lower radial oscillation speeds. According to the central limit theorem, the distribution of the average radial oscillation speed at each of the five scales is computed and assumed to be Gaussian for both normal and abnormal spirals (Bishop, 1995; Lacquaniti et al., 1987). This is performed by modeling a class mean and a class variation (#s , " s ) for each Author's personal copy H. Wang et al. / Journal of Neuroscience Methods 171 (2008) 264–270 267 Fig. 2. Row 1, a sample computer-generated spiral used in our test. Row 2, spiral drawn by a normal control. Rows 3–5 are representative spirals drawn by patients with increasingly abnormal spiral execution. Column 1: spirals with the detected local optimal centers, where the x-axis and y-axis are labeled in pixels. Column 2: polar expressions with first non-zero pressure points taken as the spiral centers (polar expression with initial center for the synthesized spiral). Column 3: polar expressions with locally optimal centers. The fitted curves are also illustrated in columns 2 and 3. For columns 2 and 3, the x-axis is labeled in radians, the y-axis is labeled in pixels. Author's personal copy 268 H. Wang et al. / Journal of Neuroscience Methods 171 (2008) 264–270 Fig. 3. Left: the multi-scale representation for a sample spiral from finer to coarser scales. The spiral is displayed in pixels at 0.1 mm/pixel, the resolution of the digitizing tablet. The corresponding radial oscillation speed traces (right) contain high frequency signals sequentially removed from one scale to the next (see text for details). x-Axis: spiral points. y-axis: unit in 1 mm/s. A level radial oscillation speed (horizontal line) is shown in the right traces for comparison cross-different scales, and corresponds to a constant power at different frequency domains. Note that tremors in this example are present only in the first three scales. Author's personal copy 269 H. Wang et al. / Journal of Neuroscience Methods 171 (2008) 264–270 scale which are estimated from training data. Given two estimated classes (#s1 , "1s ), (#s2 , "2s ) and a testing spiral with measured average radial oscillation speed xs , classification was performed at each scale by evaluating the following probability ratio: rs = (1/"1s ) exp[−(xs − #s1 )2 /2"1s2 ] (2) (1/"2s ) exp[−(xs − #s2 )2 /2"2s2 ] If r > 1, the probability that the new spiral belonged to the first class was greater than the probability that it belonged to the second class. We perform pair-wise classification tests between NC against patients with movement disorders, classifying NC vs. PD, NC vs. ET and NC vs. DY at each of the five scales separately. To avoid the bias introduced by partitioning the data into training and testing sets, we apply the “leave one out” cross-validation strategy (Bishop, 1995) for each classification test. At each iteration, one spiral from each of the paired testing classes is chosen for testing and the remaining spirals are used as training data. This training and classification is repeated until every combination of the spiral pair is selected for testing. We need to perform 100 × 100 or 10,000 training and classification iterations for each classification experiment. We measure the classification performance at each of the five scales to obtain profiles in each frequency domain. We also measure the classification performance using the aggregate of the five scales to determine the overall classification performance using all information (Table 1). 3. Results From the computer-generated spirals, the average center detection errors using 1st, 2nd, and 3rd order polynomial curve fitting are 10.91, 10.78, and 10.58 pixels, respectively. The number of found local optimal centers with error larger than 20 pixels, ∼5% of the largest radius, is 2, 2, and 2, respectively. It is evident that this approach shows robustness against noise and initialization. Furthermore, we did not observe large difference using different order polynomials for curve fitting. For the remainder of the testing, we use only 3rd order polynomial fitting. Fig. 2 gives an example of the synthesized spiral used in this test. Fig. 2 illustrates sampled data using our algorithm for improved spiral center detection in a series of actual drawn spirals of increasing abnormality. Note that the optimal center location behaves as a filter that removed the low frequency sinusoidal curve from the polar coordinates caused by an inaccurate center. High frequency tremors and other abnormal movements, however, are not affected. Employing the Gaussian pyramid technique, we separate and quantify superimposed tremors and other complex movements. Fig. 3 shows an example Gaussian pyramid series for a sample patient spiral. Since smoothing with Gaussian kernel is equivalent to low-pass filtering, the spiral at the next coarser scale only contains the low frequency signal of the previous scale. The difference between two consecutive scales reflects the remaining high frequency signals. Hence, multi-scale analysis decomposes spirals into different frequency domains. Fig. 4. The mean radial oscillation speeds of the spiral data measured with predefined spiral centers for normal controls and patients (NCpre, ETpre, DYpre, PDpre), and with optimally determined spiral centers (NCop, ETop, DYop, PDop). x-Axis: scale index for multi-scale analysis. y-Axis: the average radial oscillation speed (mm/s). Our optimal center detection did not significantly affect the classification performance at the1st and 2nd scales, which indicates that using the local optimal centers does not affect high frequency spiral information. However, at the 3rd, 4th and 5th scales, our method significantly improved the classification performance as shown by average correct classification rates in Table 1. For DY and ET, the improvements are notably the greatest at 7% and 12%, respectively. For further demonstration, the average radial oscillation speed at each of the five scales is represented for each spiral. A direct comparison of the mean average radial oscillation speeds for each clinical group – based on spirals with centers detected by previous methods and by spirals with optimal center detection – is illustrated in Fig. 4. 4. Discussion Computerized spiral analysis is used as a clinical test to record Archimedean spirals drawn on a digitizing tablet (Pullman, 1998). While highly successful in quantifying aspects of upper limb motor control (Elble et al., 1990; Louis et al., 2006) and in diagnosing and detecting early changes in movement disorders such as Parkinson’s disease (Saunders-Pullman et al., 2008), defining the center of a drawn spiral can be problematic. In this report, we improve on the detection of the center of the Archimedean spiral based on regression analysis. With the new algorithms, the locally optimal spiral center was shown to result in a more regular polar transformation. We found that accurately determining the center of the spiral subsequently reduces spurious low frequency noise and improves the overall utility of spiral analysis. Applying the new center selection process to spirals drawn by normal controls and patients with movement disorders resulted in Table 1 Classification results list by percentage correct 1st scale NC: optimal NC: predefined 2nd scale 3rd scale 4th scale 5th scale All five scales PD ET DY PD ET DY PD ET DY PD ET DY PD ET DY PD ET DY 97 98 98 98 84 79 97 99 100 100 88 88 85 82 99 99 88 83 74 68 88 83 81 76 57 58 76 64 73 66 98 98 100 99 90 89 Multi-scale results comparing control subjects to each patient group using optimal and pre-defined centers. NC = normal control; PD = Parkinson’s disease; ET = essential tremor; DY = dystonia. Author's personal copy 270 H. Wang et al. / Journal of Neuroscience Methods 171 (2008) 264–270 modest but improved classification rates in low frequency domains. We found that the improvement is as high as 12%, which is notable considering the clinical subtleties and overlapping kinematics of movement disorders (Deuschl et al., 2001). This result demonstrates that our local optimal spiral centers can significantly improve the accuracy of measured spiral indices at low frequency domains by removing artifacts caused by inaccurate spiral centers. Spiral centers calculated with the new algorithm revealed greater analytical accuracy, improved clinical classification rates and decreased false positive tremor assessments compared to spiral analysis without optimal center detection. The center detection algorithm, combined with the Gaussian pyramid and multi-scale spiral representations, enhanced the clinical utility of spiral analysis. These new methods thus offer an improvement in spiral analysis measurements and diagnostic utility in assessing clinical motor performance. References Bishop CM. Neural networks for pattern recognition. Oxford New York: Clarendon Press Oxford University Press; 1995. Deuschl G, Raethjen J, Lindemann M, Krack P. The pathophysiology of tremor. Muscle Nerve 2001;24:716–35. Elble RJ, Sinha R, Higgins C. Quantification of tremor with a digitizing tablet. J Neurosci Methods 1990;32:193–8. Elble RJ, Brilliant M, Leffler K, Higgins C. Quantification of essential tremor in writing and drawing. Mov Disord 1996;11:70–8. Forsyth D, Ponce J. Computer vision: a modern approach. Upper Saddle River, NJ: Prentice Hall; 2003. Hogg RV, McKean JW, Craig AT. Introduction to mathematical statistics. 6th ed. Upper Saddle River, NJ: Pearson Education; 2005. Lacquaniti F, Ferrigno G, Pedotti A, Soechting JF, Terzuolo C. Changes in spatial scale in drawing and handwriting: kinematic contributions by proximal and distal joints. J Neurosci 1987;7:819–28. Lindeberg T. Scale-space theory in computer vision. Kluwer Academic Publishers; 1994. Liu X, Carroll CB, Wang SY, Zajicek J, Bain PG. Quantifying drug-induced dyskinesias in the arms using digitised spiral-drawing tasks. J Neurosci Methods 2005;144:47–52. Louis ED, Yu Q, Floyd AG, Moskowitz C, Pullman SL. Axis is a feature of handwritten spirals in essential tremor. Mov Disord 2006;21:1294–5. Miralles F, Tarongi S, Espino A. Quantification of the drawing of an Archimedes spiral through the analysis of its digitized picture. J Neurosci Methods 2006; 152:18–31. Pullman SL. Spiral analysis: a new technique for measuring tremor with a digitizing tablet. Mov Disord 1998;13(Suppl. 3):85–9. Saunders-Pullman R, Derby C, Stanley K, Floyd A, Bressman S, Lipton RB, et al. Validity of spiral analysis in early Parkinson’s disease. Mov Disord 2008;23: 531–7. Siebner HR, Ceballos-Baumann A, Standhardt H, Auer C, Conrad B, Alesch F. Changes in handwriting resulting from bilateral high-frequency stimulation of the subthalamic nucleus in Parkinson’s disease. Mov Disord 1999;14: 964–71. Wang S, Bain PG, Aziz TZ, Liu X. The direction of oscillation in spiral drawings can be used to differentiate distal and proximal arm tremor. Neurosci Lett 2005;384:188–92. Yu SX. Segmentation using multiscale cues. In: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 1; 2004, I-247–I-54.