Spiral drawing: Quantitative analysis and artificial-intelligence-based diagnosis using a smartphone

Journal of the Neurological Sciences 411 (2020) 116723

Contents lists available at ScienceDirect

Journal of the Neurological Sciences journal homepage: www.elsevier.com/locate/jns

Spiral drawing: Quantitative analysis and artificial-intelligence-based diagnosis using a smartphone Nobuyuki Ishiia, Yuki Mochizukib, Kazutaka Shiomia, Masamitsu Nakazatoa, Hitoshi Mochizukia, a b

T ⁎

Division of Neurology, Respirology, Endocrinology and Metabolism, Department of Internal Medicine, Faculty of Medicine, University of Miyazaki, Miyazaki, Japan Department of Agricultural and Environmental Sciences, Faculty of Agriculture, University of Miyazaki, Miyazaki, Japan

A R T I C LE I N FO

A B S T R A C T

Keywords: Artificial intelligence Tremor Smartphone Machine learning Spiral Cerebellar ataxia

Background: The evaluation of neurological examination in clinical practice still remains qualitative or semiquantitative, and the results often vary depending on an examiner's skill level and are less objective. In this study, we developed a smartphone-based application to investigate quantifying neurological examinations using hand-drawn spirals and diagnose patients with tremor using artificial intelligence (AI). Methods: This study included 24 and 26 patients with essential tremor (ET) and cerebellar disease (CD), respectively, and 41 age-matched normal controls (NCs). We obtained 69, 46, and 56 hand-drawn spirals from the NC, ET, and CD groups, respectively, as image data captured by smartphones. The patients traced a printed reference spiral. The length of this spiral was compared with the reference spiral length (% of spiral length) and the total deviation area between these spirals was calculated. The server also estimates the diagnostic probability through AI. Results: The quantified spiral analysis (% of spiral length and deviation area) significantly correlated with disease severity in each disease group, and significant differences in the deviation area were observed among all groups. The AI diagnosis showed 79%, 70%, and 73% accuracies for the NC, ET, and CD groups, respectively. Conclusion: This study indicates the possibility of using a smartphone as a medical examination tool and demonstrates the application of AI in neurological examinations.

1. Introduction

Human life has greatly improved with the development in technology. In the medical field, artificial intelligence (AI)-based diagnostic systems have been introduced mainly in radiological and pathological areas [4–6]. In this study, we developed a smartphone-based application with AI to analyze the hand-drawn spirals by patients and verified the effectiveness of the AI diagnosis. This technology may help patients with tremor decide whether they need to visit doctors, and help general physicians decide whether they must refer to neurologists. This is the first study to analyze neurological examinations using AI.

Neurological examination in clinical practice is almost qualitative or semi-quantitative. In the evaluation of muscle weakness, except handgrip strength, for example, we evaluate muscle-power grading according to the Medical Research Council (MRC) scale, which ranges from 0 to 5 [1]. Tremor severity is graded through several scales, such as Fahn–Tolosa–Marin tremor rating scale [22] and/or the essential tremor rating assessment scale (TETRAS) [23], and cerebellar ataxia is often evaluated using the International Cooperative Ataxia rating scale (ICARS) [2] and/or the Scale for the Assessment and Rating of Ataxia (SARA) [3]. These semi-quantitative scales are influenced by the examiner; thus, the results often vary depending on each examiner's skill level, and the results are less objective. One of the purposes of this study is to make qualitative neurological examinations quantitative. We developed a simple quantitative method for diagnosing tremor and cerebellar ataxia by using a smartphone, which is already used worldwide.

2. Materials and methods 2.1. Subjects During the period from April 2012 to June 2019, patients with essential tremors (ET; n = 24), hereditary spinocerebellar ataxia [SCA, n = 14; genetically confirmed SCA3 (n = 1) and SCA6 (n = 12) and

⁎

Corresponding author at: Division of Neurology, Respirology, Endocrinology and Metabolism, Department of Internal Medicine, University of Miyazaki, 5200 Kihara, Kiyotake, Miyazaki 889-1692, Japan. E-mail addresses: [email protected] (N. Ishii), [email protected] (K. Shiomi), [email protected] (M. Nakazato), [email protected] (H. Mochizuki). https://doi.org/10.1016/j.jns.2020.116723 Received 10 December 2019; Received in revised form 1 February 2020; Accepted 3 February 2020 Available online 04 February 2020 0022-510X/ © 2020 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/BY-NC-ND/4.0/).

Journal of the Neurological Sciences 411 (2020) 116723

N. Ishii, et al.

genetically not confirmed, pure cerebellar SCA (n = 1)], or multiple system atrophy (cerebellar type without rigidity nor myoclonus, MSAC; n = 12) who visited the outpatient department or were hospitalized at the University of Miyazaki Hospital and who were asked to draw spirals during their examinations participated in this study. Their diagnoses (ET and MSA-C) were based on the diagnostic criteria [7,8]. Patients with SCA or MSA-C were defined as the groups with cerebellar disease (CD). Two neurologists confirmed the diagnosed and evaluated the severity of the patients' tremor based on #6 (Archimedes Spirals) using the Performance Subscale feature in TETRAS [23] for ET and #14 (Drawing of the Archimedes spiral on a predrawn pattern) of ICARS [2] for CD. Both scales range from 0 to 4, where 0 indicates normal function and 4 is the lowest level of function. Forty-one age-matched normal controls (NCs) were also enrolled in this study. Each subject handwrote 1–3 spirals. The protocol was approved by the Ethics Committee of the University of Miyazaki, with a waiver of written informed consent obtained from all participants. 2.2. Data acquisition and analysis Actual experiments are completed in the following three steps. First, the reference spirals are printed on an A4-sized paper. Second, the subject traces the spiral from outside. Third, we take a photo of their tracing with a smartphone installed with the analysis software. The results are automatically displayed on the smartphone. 2.3. Reference spiral on paper The reference spiral is defined as follows:

t ∈ {100, 101, 102,….1960}

Fig. 1. (A) A reference spiral. Subjects traced the reference spiral with a red pen from the outside. Spiral width is approximately 100 mm and distance between each marker printed at the four corners is 171 mm. Typical hand-drawn spiral of (B) normal controls (NC), (C) patients with essential tremor (ET), and (D) patients with cerebellar disease (CD). The spiral-length percentage and deviation areas of ET and CD are larger than those of NC (spiral length of NC, ET, and CD are 100.9%, 115.7%, and 118.5%, respectively; corresponding deviation areas are 558, 1356, and 1551 mm2). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

t ’ = √ (t /4) x (t ) = cos(t ’) − t ’∙sin(t ’) y (t ) = sin(t ’) + t ’∙cos(t ’) We drew a reference spiral with dotted lines and printed it so that the width on the x-axis is approximately 100 mm. Markers were printed at the four corners away from the spiral image (indicated in Fig. 1A), and these were used as reference positions in later calculations.

distance between the adjacent reference points was calculated, and their sum is defined as the “deviation area”.

2.4. Process on smartphone and server The subject traced the printed reference spiral with a red pen from the outside. The traced paper was photographed with a smartphone, and data were sent to the server. Based on the image data, the server calculates the total length of the hand-drawn spiral and the total deviation value between the reference and hand-drawn spirals. The server also estimates the diagnostic probability through AI.

2.6. AI as a diagnostic tool The convolutional neural network (CNN) was originally described by LeCun [9] and is one of the most popular neural network algorithms for pattern recognition. The CNN operation is particularly advantageous in finding local patterns, such as edges, sharps, lines, or other visual elements in images [10]. In this study, the authors used the popular and standard CNN method [11], as indicated in Fig. 2. The image data sent from the smartphone were resized to 128 × 128 pixels, and the hand-drawn thick line was treated as a thin line with the middle points of the line width as a representative. Multiple filters were applied on to the two convolutional layers to generate multiple feature maps. In addition to the convolution layers, a max-pooling layer was used to perform reduction, and a softmax layer was used to perform classification. Rectified linear unit (ReLu) [12] is conventionally used as an activation function for the neural networks. For modeling categorical probability distributions, we used softmax as an output activation function [13]. Classification was performed through probability distributions (percentage) so that the total is 100%. From 171 hand-drawn images (69, 46, and 56 for NC, ET, and CD, respectively), 139 images (55, 38, and 46 of NC, ET, and CD, respectively) were randomly selected as a dataset and used as machine-

2.5. Quantitative evaluation By deforming the hand-drawn image by using the markers at the four corners, the reference spiral line can be accurately recognized from the relative position to the marker. Next, for all points on the reference spiral, we identified the point on the red hand-drawn line closest to the reference spiral on the normal vector. As such, points constituting 1860 hand-drawn lines were obtained. The total length of the hand-drawn spiral is the sum of the distances between adjacent points to the points that make up 1860 hand-drawn lines. We define the percentage of spiral length as (length of the handdrawn line/length of the reference spiral line) × 100. The total deviation area was calculated as follows. For the points composing the 1860 hand-drawn lines, we measured the distance between the point on the reference spiral and the corresponding hand-drawn line composing the point. The value obtained by multiplying the measured value and the 2

Journal of the Neurological Sciences 411 (2020) 116723

N. Ishii, et al.

Fig. 2. Convolutional neural network formula used in this study. The image data sent from the smartphone are resized to 128 × 128 pixels. Two convolutional layers were applied multiple filters to generate multiple feature maps. In addition to the convolution layers, a max-pooling layer was used to perform reduction, and the softmax layer performs classification. Classification was performed based on probability distributions (percentage) amounting to 100%.

learning data. The CNN system was trained using this dataset. The remaining 33 images (13, 9, and 11 for NC, ET, and CD, respectively) were classified into three groups (NC, ET, or CD) by using the trained CNN. The dataset for machine learning was randomly selected five times, and the remaining 33 images were classified using CNN.

Table 1 Patients' characteristics. Normal controls

Essential tremor

Cerebellar disease

41 67.4 (10.9) 19 (46.3)

24 73.0 (12.4)

26 66.1 (12.4)

15 (62.5) 2 (8.3)

11 (42.3) 7 (26.9)

F = 2.52, p = .086 p = .303 p = .344

11 (45.8) 9 (37.5) 2 (8.3) 107.1 (8.2)

9 (34.6) 7 (26.9) 3 (11.5) 111.9 (16.4)

p < .001

1058.6 (275.5)

1296.6 (504.5)

F = 34.82, p < .001

1

46 4 (8.7)

56 13 (23.2)

p = .141

2 3 4

23 (50.0) 15 (32.6) 4 (8.7)

23 (41.1) 12 (21.4) 8 (14.3)

Patients No. Age (SD)

2.7. Statistical analysis

Men (%) Severity scale (%)

We performed statistical analysis for patient's characteristics and quantitative tremor severity (%spiral length and deviation area) among the three groups (NC, ET, and CD) using the following methods: for continuous variables, we used the Shapiro–Wilk normality test to check whether the parametric model was appropriate; we then used analysis of variance (ANOVA) or Kruskal–Wallis analysis, where appropriate, followed by post-hoc Holm's test. To analyze categorical variables, we used Fisher's exact test. The association between clinical severity scale and the quantitative tremor severity was evaluated using the Jonckheere–Terpstra trend test. Receiver operating characteristic (ROC) analyses and the areas under the curves (AUC) in ROC analyses were conducted to evaluate the diagnostic performance of the CNN. Statistical analyses were conducted using the R software (version 3.5.1; R Foundation for Statistical Computing, Vienna, Austria), and P < .05 was considered statistically significant.

1 2 3 4

%Spiral length (SD) Deviation area, mm2 (SD) Spirals No. Severity scale (%)

101.6 (1.1) 725.6 (166.3) 69

Statistic value

SD, standard deviation; severity scales were #6 (Archimedes Spiral) in Performance Subscale in The Essential Tremor Rating Assessment Scale (TETRAS) and #15 (Drawing of Archimedes' spiral on a predrawn pattern) of International Cooperative Ataxia Rating Scale in essential tremor and cerebellar disease, respectively. Age and Deviation area, and %Spiral length are analyzed with analysis of variance (ANOVA) or Kruskal–Wallis analysis, respectively. Gender and Severity scale are compared using Fisher's exact test.

3. Results 3.1. Patients' characteristics Table 1 summarizes the patients' characteristics in this study. Fortyone NCs (age 67.4 ± 10.9 years), 24 patients with ET (age 73.0 ± 12.4 years), and 26 patients with CD (66.1 ± 12.4 years) participated in this study. The three groups showed no significant differences with respect to age and sex. A total of 69, 46, and 56 handdrawn spirals were obtained from NCs and patients with ET and CD, respectively.

percentage and deviation area were significantly positively correlated with disease severity (Fig. 3C–D).

3.3. AI diagnosis We defined “AI diagnosis” as the largest probability distribution diagnosis among NC, ET, or CD and compared it with actual diagnosis. Table 2 demonstrates the sensitivity, specificity, and accuracy of the AI diagnosis. The specificity in NC and CD were more than 80%, although the sensitivity of the three groups was less than 65%. Accuracy showed 79%, 70%, and 73% for the NC, ET, and CD groups, respectively. Fig. 4 shows the ROC analysis for the probability distributions of AI diagnosis. The AUC and their 95% confidence interval (95% CI) of NC, ET, and CD were 0.88 (0.83–0.93), 0.69 (0.59–0.79), and 0.80 (0.73–0.87), respectively.

3.2. Quantification of tremor severity Figs. 2B–D show the typical hand-drawn spirals of NCs and patients with ET or CD. The spirals of ET and CD indicate large spiral-length percentage and deviation area compared with that of NC. Fig. 3A–B show the spiral-length percentage and deviation area of the three groups. The CD group shows the largest spiral length and deviation area. Regarding the spiral-length percentage, significant differences existed only between NC and CD, while the deviation area significantly differed for all three groups. In both ET and CD, the spiral-length 3

Journal of the Neurological Sciences 411 (2020) 116723

N. Ishii, et al.

A

B

p < 0.001 p < 0.001

Fig. 3. Distributions of the percentages of spiral lengths (%Spiral Length) and deviation areas of NC, ET, and CD groups. (A) Spirals of normal controls and CD or ET patients show significant difference in spiral–length percentage. (B) Significant difference is observed among the three groups with respect to deviation areas. Correlation of the scale of disease severity in ET and CD with (C) %spiral length and (D) deviation area, all of which were significantly positively correlated with disease severity (p for trend < 0.001). Each box plot represents the median, interquartile range (IQR), minimum (first quartile −1.5 × IQR), and maximum values (third quartile +1.5 × IQR). NC, normal control; ET, essential tremor; CD, cerebellar disease.

p < 0.001

p = 0.260

p < 0.001

p = 0.037

Deviation Area (mm2)

%Spiral Length (%)

180

160

140

120

2000

1000

100

NC

%Spiral Length (%)

C

180

CD

ET

NC

CD

ET

CD For trend, p < 0.001

ET For trend, p < 0.001

160

140

120

100 1

Deviation Area (mm2)

D

3000 2500

3

2

4

Severity

1

2

3

4

3

4

CD For trend, p < 0.001

ET For trend, p < 0.001

2000 1500 1000

1

3

2

4

Severity

1

2

ET or CD) with machine learning of the hand-drawn spiral images showed 85% specificity and 79% accuracy.

Table 2 Sensitivity, specificity, and accuracy of artificial intelligence (AI) diagnosis. Normal control (%)

Essential tremor (%)

Cerebellar disease (%)

Median

[max, min]

Median

[max, min]

Median

[max, min]

62 85 79

[92, 54] [95, 65] [85, 70]

44 79 70

[89, 22] [88, 54] [88, 52]

64 91 73

[73, 27] [95, 59] [85, 64]

4.1. Quantitative analysis of hand-drawn spirals

Sensitivity Specificity Accuracy

The calculated quantitative tremor severity strongly correlates with clinical severity scale, suggesting that the proposed method reflects clinical symptoms accurately. Furthermore, this smartphone-based spiral analysis could calculate the length and deviation area of the hand-drawn spiral through a built-in camera, by which we could assess the changes of symptom severity quantitatively. Most of the previous studies on digital spiral analysis were conducted using a digital tablet [14–16]. These digital tablet-based analyses obtained 4-dimensional information of spiral writing such as spiral shape (X- and Y-axes), pen pressure (Z-axis), and writing speed. These data were analyzed mathematically, such as by using fast Fourier transform, thus resulting in several parameters including mean pressure, pressure power, pressure frequency, drawing smoothness, and

4. Discussion In this study, we demonstrated the following. (1) The quantitative tremor severity calculated through our system strongly correlated with clinical severity scale. (2) The quantitative hand-drawn spiral analysis conducted using the smartphone-based application could differentiate between NCs and patients with ET or CD. (3) AI-based diagnosis (NC vs. 4

Journal of the Neurological Sciences 411 (2020) 116723

N. Ishii, et al.

4.3. Clinical implication The use of a smartphone-based application has four advantages in clinical practice. First, in contrast to previous studies [14–16] that used tablets and PCs, our system consists only of the use of a smartphone and its built-in camera. This implies its easy distribution to general physicians and ease in updates. Furthermore, additional equipment is not needed for its preparation. Second, quantification of the tremor's characteristics may help in the objective evaluation of its treatment. Third, by using this application, people with shaking limbs could decide whether they should visit a doctor, and general physicians could decide whether they should refer patients with tremors to a neurologist. Fourth, this application has a high expandability to add other diseases in the AI diagnosis because we used simple spiral images for machine learning. Thus, in the future, it may be possible to differentiate other tremorigen syndromes such as Parkinson's disease, dystonia, and patients with scans without evidence of dopaminergic deficits (SWEDDs) [21]. 4.4. Study limitations

Fig. 4. Receiver operating characteristic (ROC) curves of normal controls (NC), essential tremor (ET), and cerebellar disease (CD). The area under curves (AUCs) and their 95% confidence interval (95% CI) of NC, ET, and CD are 0.88 (0.83–0.93), 0.69 (0.59–0.79), and 0.80 (0.73–0.87), respectively.

This study has certain limitations. First, the patients with tremor in this study included only those with ET or CD. Patients with CD included SCA and MSA-C; these have different components of tremor even if showing only cerebellar dysfunction in neurological examination. Moreover, people with physiologic tremors and tremors, such as Parkinson's disease or dystonia, were not enrolled. Second, results could have been affected by the quality of the paper and pen. As such, the normal range should be set carefully. Third, because the accuracy of AI diagnosis remains at a low level of 70–80%, AI diagnosis is just one of assistant tools in clinical practice.

index of the spiral loop-to-loop width variability [14,16]. Such a system is not simple and cannot easily be applied to clinical practices. In contrast to these previous studies, less information, such as the two-dimensional spiral shape, was obtained in the current study. This resulted in fewer data on tremor quality; however, this simple system proved sufficient for quantifying tremor severity and differentiating whether patients' tremor was pathological. Our application on a smartphone with a built-in camera is easy-to-use and does not require the preparation of additional equipment such as tablets and PCs. This application could not detect significant differences in %spiral length between ET and CD. Conducting a post hoc power analysis based on the statistical power of 0.8, 104 patients were required in each group. Thus, no significant differences in %spiral length might be caused by a small sample size.

5. Conclusion We demonstrated that our smartphone-based application could quantify tremor severity though hand-drawn spirals and would help in the AI-based diagnosis of patients with tremors. This study indicates the possibility of using a smartphone as a medical examination tool, and demonstrates the application of AI in neurological examinations. Declaration of Competing Interest

4.2. AI for tremor diagnosis

None.

In this study, our application showed 79% accuracy in differentiating NC vs. non-NC (ET or CD) in patients with tremor. No previous studies have diagnosed tremor by using AI with hand-drawn spirals; thus, it is unknown whether the obtained accuracy is high or low. Recently, AI-based diagnosis in neurology has spread significantly, especially in imaging studies [17]. For example, Kiryu et al. [18] demonstrated AI-based diagnosis for Parkinsonism, such as Parkinson's disease, progressive supranuclear palsy, and multiple system atrophy, by using brain MRIs, with accuracy of almost 95%. One of the reasons for the lower accuracy of our study in comparison could be the low sensitivity (44%) of the diagnosis of ET. Through the five cross-validation tests, 16% and 27% of the ET spirals were misdiagnosed as NC and CD, respectively. The spirals of ET patients with low severity (mean = 1.4) were judged as NC, whereas the severity of ET patients mistaken for CD (mean = 2.6) was almost the same as the total ET severity (mean = 2.3). We consider the low sensitivity of ET for the following reasons: (1) slight tremors in NCs, including physiologic or senile tremor may generally be difficult to differentiate from mild ET [19], and (2) the pathophysiology of ET contains elements of cerebellar dysfunctions [20]. These problems might be solved by learning more spirals and using patients' age as a parameter in the application.

Acknowledgement This study was partly supported by JSPS KAKENHI Grant Number JP19K20715 (NI) and JP18K10052 (HM). References [1] Medical Research Council, Aids to the Examination of the Peripheral Nervous System, HM Stationery Office, London, 1976. [2] P. Trouillas, T. Takayanagi, M. Hallett, R.D. Currier, S.H. Subramony, K. Wessel, A. Bryer, H.C. Diener, S. Massaquoi, C.M. Gomez, P. Coutinho, M.B. Hamida, G. Campanella, A. Filla, L. Schut, D. Timann, J. Honnorat, N. Nighoghossian, B. Manyam, International cooperative Ataxia rating scale for pharmacological assessment of the cerebellar syndrome. The Ataxia neuropharmacology committee of the World Federation of neurology, J. Neurol. Sci. 145 (2) (1997) 205–211, https:// doi.org/10.1016/s0022-510x(96)00231-6. [3] T. Schmitz-Hubsch, S.T. du Montcel, L. Baliko, J. Berciano, S. Boesch, C. Depondt, P. Giunti, C. Globas, J. Infante, J.S. Kang, B. Kremer, C. Mariotti, B. Melegh, M. Pandolfo, M. Rakowicz, P. Ribai, R. Rola, L. Schols, S. Szymanski, B.P. van de Warrenburg, A. Durr, T. Klockgether, R. Fancellu, Scale for the assessment and rating of ataxia: development of a new clinical scale, Neurology 66 (11) (2006) 1717–1720, https://doi.org/10.1212/01.wnl.0000219042.60538.92. [4] D.A. Bluemke, Radiology in 2018: are you working with AI or being replaced by AI? Radiology 287 (2) (2018) 365–366, https://doi.org/10.1148/radiol.2018184007. [5] S. Jha, E.J. Topol, Adapting to artificial intelligence: radiologists and pathologists as information specialists, JAMA 316 (22) (2016) 2353–2354, https://doi.org/10.

5

Journal of the Neurological Sciences 411 (2020) 116723

N. Ishii, et al.

1001/jama.2016.17438. [6] D.D. Miller, E.W. Brown, Artificial intelligence in medical practice: the question to the answer? Am. J. Med. 131 (2) (2018) 129–133, https://doi.org/10.1016/j. amjmed.2017.10.035. [7] G. Deuschl, P. Bain, M. Brin, Consensus statement of the Movement Disorder Society on tremor. Ad Hoc scientific committee, Mov Disord 13 (Suppl. 3) (1998) 2–23, https://doi.org/10.1002/mds.870131303. [8] S. Gilman, G.K. Wenning, P.A. Low, D.J. Brooks, C.J. Mathias, J.Q. Trojanowski, N.W. Wood, C. Colosimo, A. Durr, C.J. Fowler, H. Kaufmann, T. Klockgether, A. Lees, W. Poewe, N. Quinn, T. Revesz, D. Robertson, P. Sandroni, K. Seppi, M. Vidailhet, Second consensus statement on the diagnosis of multiple system atrophy, Neurology 71 (9) (2008) 670–676, https://doi.org/10.1212/01.wnl. 0000324625.00404.15. [9] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document recognition, Proc. IEEE 86 (11) (1998) 2278–2324. [10] A. Arafati, P. Hu, J.P. Finn, C. Rickers, A.L. Cheng, H. Jafarkhani, A. Kheradvar, Artificial intelligence in pediatric and adult congenital cardiac MRI: an unmet clinical need, Cardiovasc Diagn Ther 9 (Suppl. 2) (2019), https://doi.org/10. 21037/cdt.2019.06.09 S310-S25. [11] E. Urtnasan, J.U. Park, E.Y. Joo, K.J. Lee, Automated detection of obstructive sleep apnea events from a single-lead electrocardiogram using a convolutional neural network, J. Med. Syst. 42 (6) (2018) 104, , https://doi.org/10.1007/s10916-0180963-0. [12] R.H. Hahnloser, R. Sarpeshkar, M.A. Mahowald, R.J. Douglas, H.S. Seung, Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit, Nature 405 (6789) (2000) 947–951, https://doi.org/10.1038/35016072. [13] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT press, 2016. [14] J. Ratliff, R.A. Ortega, H.Y. Ooi, A. Mirallave, A. Glickman, Q. Yu, D. Raymond, S. Bressman, S. Pullman, R. Saunders-Pullman, Digitized spiral analysis may be a potential biomarker for brachial dystonia, Parkinsonism Relat. Disord. 57 (2018) 16–21, https://doi.org/10.1016/j.parkreldis.2018.07.004. [15] A.P. Legrand, I. Rivals, A. Richard, E. Apartis, E. Roze, M. Vidailhet, S. Meunier,

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

6

E. Hainque, New insight in spiral drawing analysis methods - application to action tremor quantification, Clin. Neurophysiol. 128 (10) (2017) 1823–1834, https://doi. org/10.1016/j.clinph.2017.07.002. M. San Luciano, C. Wang, R.A. Ortega, Q. Yu, S. Boschung, J. Soto-Valencia, S.B. Bressman, R.B. Lipton, S. Pullman, R. Saunders-Pullman, Digitized spiral drawing: a possible biomarker for early Parkinson’s disease, PLoS One 11 (10) (2016) e0162799, , https://doi.org/10.1371/journal.pone.0162799. R. Bruffaerts, Machine learning in neurology: what neurologists can learn from machines and vice versa, J. Neurol. 265 (11) (2018) 2745–2748, https://doi.org/ 10.1007/s00415-018-8990-9. S. Kiryu, K. Yasaka, H. Akai, Y. Nakata, Y. Sugomori, S. Hara, M. Seo, O. Abe, K. Ohtomo, Deep learning to differentiate parkinsonian disorders separately using single midsagittal MR imaging: a proof of concept study, Eur. Radiol. 29 (12) (2019) 6891–6899, https://doi.org/10.1007/s00330-019-06327-0. A. Schrag, A. Munchau, K.P. Bhatia, N.P. Quinn, C.D. Marsden, Essential tremor: an overdiagnosed condition? J. Neurol. 247 (12) (2000) 955–959, https://doi.org/10. 1007/s004150070053. R. Hanajima, R. Tsutsumi, Y. Shirota, T. Shimizu, N. Tanaka, Y. Ugawa, Cerebellar dysfunction in essential tremor, Mov. Disord. 31 (8) (2016) 1230–1234, https://doi. org/10.1002/mds.26629. T. Schirinzi, F. Di Lorenzo, V. Ponzo, M.G. Palmieri, A.R. Bentivoglio, O. Schillaci, A. Pisani, G. Koch, Mild cerebello-thalamo-cortical impairment in patients with normal dopaminergic scans (SWEDD), Parkinsonism Relat. Disord. 28 (2016) 23–28, https://doi.org/10.1016/j.parkreldis.2016.03.023. S. Fahn, E. Tolosa, C. Marin, Clinical rating scale for tremor, Parkinson’s Disease and Movement Disorder, Lippincott Williams & Wilkins, Baltimore, 1993, pp. 271–280. R. Elble, C. Comella, M. Fahn, M. Hallett, J. Jankovic, J.L. Juncos, P. Lewitt, K. Lyons, W. Ondo, R. Pahwa, K. Sethi, N. Stover, D. Tarsy, C. Testa, R. Tintner, R. Watts, T. Zesiewicz, Reliability of a new scale for essential tremor, Mov. Disord. 27 (2012) 1567–1569, https://doi.org/10.1002/mds.25162.