Suitability of automated writing measures for clinical trial outcome in writer’s cramp disease

Abstract

Background:

Writer’s cramp (WC) dystonia is a rare disease that causes abnormal postures during the writing task. Successful research studies for WC and other forms of dystonia are contingent on identifying sensitive and specific measures that relate to the clinical syndrome and achieve a realistic sample size to power research studies for a rare disease. Although prior studies have employed writing kinematics, their diagnostic performance remains unclear. This study aimed to evaluate the diagnostic performance of automated measures that distinguish WC from healthy.

Methods:

21 WC and 22 healthy subjects performed a sentence-copying assessment on a digital tablet using kinematic and hand recognition softwares. The sensitivity and specificity of automated measures were calculated using a logistic regression model. Power analysis was performed for two clinical research designs using these measures. The test and retest reliability of select automated measures was compared across repeat sentence-copying assessments. Lastly, a correlational analysis with subject-rated and clinician-rated outcomes was performed to understand the clinical meaning of automated measures.

Results:

Of the 23 measures analyzed, the measures of word legibility and peak accelerations distinguished WC from healthy with high sensitivity and specificity, demonstrated smaller sample sizes suitable for rare disease studies, and the kinematic measures showed high reliability across repeat visits while both word legibility and peak accelerations measures showed significant correlations with the subject and clinician-rated outcomes.

Discussion:

Novel automated measures which capture key aspects of the disease and are suitable for use in clinical research studies of WC dystonia were identified.

Graphical Abstract

Identification of automated outcome measures is a pre-requisite to well-designed clinical research studies testing therapies. This study evaluated the diagnostic performance of automated writing measures that distinguish writer’s cramp from healthy. Four automated writing measures show high sensitivity and specificity to predict WC status.

INTRODUCTION

Dystonia is an involuntary movement disorder characterized by sustained or intermittent muscle contractions that lead to abnormal postures (1–3). Dystonia can be generalized in the whole body or focal to a single body part, with focal dystonias occurring more commonly (4–6). Task-specific focal dystonias occur during a specific motor task. Writer’s cramp (WC) is an example of task-specific focal hand dystonia that occurs during writing (7). WC patients report reduced quality of life due to debilitating hand pain and inability to communicate via writing (7). Over time, a subset of WC patients also develop difficulty with other fine motor tasks (7). There are no disease-modifying therapies available for dystonia, and symptomatic treatments in the form of botulinum toxin injections and oral medications provide limited efficacy (2). Clinical research studies to identify new targets for therapy are much needed; however, several challenges exist for research studies in this rare movement disorder.

With an estimated prevalence of 15 per million persons, WC is a rare disorder that precludes large randomized placebo-controlled design for research studies (2,4–6). Another challenge is that the current outcome measures for evaluating the efficacy of an intervention consist of clinician-rated outcomes such as the Burke Fahn Marsden (BFM) dystonia scale and writer’s cramp rating scale (WCRS) (8,9). Clinician-rated outcomes require specialized practitioners and are susceptible to inter-rater variability. In addition, the coarseness of categorical gradations in clinician-rated outcomes might limit the ability to detect a differential response to research manipulations. Lastly, while subject-rated outcomes such as the BFM disability scale and Arms Dystonia and Disability scale (ADDS) (8,10,11) are essential for quality-of-life measurements, they are not suitable as primary outcome measures in an investigational research study because they are highly susceptible to a placebo effect.

Automated motor performance measures offer an opportunity to bypass these limitations. Previous evaluations of writing kinematics using a digital tablet have reported several writing measures that can distinguish WC from healthy volunteers (HV): mean duration of writing, mean axial pressure, mean stroke frequency, normalized jerk, and coefficient of variation of peak velocity (10–16). Two of these studies additionally compared these kinematic measures to WCRS and ADDS (10,11). The automated measures did not correlate with the clinician or subject-rated outcomes, which the authors attributed to differing aspects of motor impairments captured by the writing kinematics and rated outcomes (10,11). However, the inability to correlate automated measures to the clinician or subject-rated outcomes makes it difficult to apply the research findings using these metrics to a broader effect on the clinical disease. Move over, no study to date has evaluated the sensitivity and specificity of these kinematic writing measures to differentiate WC from HV and whether such measures would have a sufficient signal-to-noise ratio to test a research hypothesis with a realistic sample size in this rare disease. Therefore, identification of automated writing measures that demonstrate high diagnostic performance are needed as a prerequisite to well-designed clinical research studies in WC.

The goal of this study was to identify automated writing measures suitable for research studies in WC subjects using a diagnostic framework. We focused on five study design features. First, we tested a broad range of automated writing measures generated by the kinematic software MovAlyzer to identify measures that could distinguish WC from HV. Since writing illegibility is a major complaint of individuals with WC (2), we also developed an automated measure of word legibility using handwriting recognition software and evaluated its ability to distinguish WC from HV. The word legibility measure is thus both an automated measure and a measure of a clinically meaningful outcome for individuals with WC. Second, we measured the diagnostic performance of the automated writing measures to distinguish WC from HV. Third, we evaluated the sample size needed to power a research study using these measures in two different clinical research designs. Fourth, we calculated these automated measures’ test and retest reliability across repeat writing assessments within subjects. Last, we compared these automated measures with their corollary subject or clinician-rated outcomes.

MATERIALS AND METHODS

Subjects:

The research study was approved by the Duke University Institutional Review Board. All isolated WC subjects with dystonia affecting the right hand were considered eligible for the research study. All enrolled subjects were more than three months from the last botulinum toxin injections and not on oral trihexyphenidyl, a symptomatic treatment for dystonia. Subjects were referred from Duke University Movement Disorders Center, where they were diagnosed by a Movement Disorder specialist with a standard medical history and neurological examination. All subjects provided written consent before participation in the research study. All 43 subjects in the study were right-hand dominant, except for one WC subject (case seven) who reported ambidexterity but wrote with his right hand. Healthy subjects who were right hand dominant without co-morbid neurological diagnoses and age-matched to a WC were recruited from Duke healthy volunteer registry and advertising in the local community.

Clinician and subject-rated outcomes:

Subjects were video recorded from the neck down (focused on their right arm) during two writing assessments. Movement Disorder clinicians (B.S. and P.T.) rated the videos using the BFM right arm dystonia scales (8) and WCRS (9) while blinded to the identity of the subjects. Clinicians were provided literature on the rating scales to establish concordant ratings. The two clinicians’ inter-rater reliability (ICC) score was 0.92 for the BFM right arm dystonia scale and 0.62 for WCRS part A Total movement score. The WCRS part A is composed of three subscore measures. The ICC of the subscore measures ranged from 0.85 to 0.64 for dystonic posture and writing tremor respectively.The ICC for WCRS part B writing speed was 0.70. Subjects also reported their disabilities using two outcome scales: BFM Disability Scale and ADDS. On both BFM and ADDS scales, handwriting disability was scored with 0 denoting normal and 4 or 3 denoting severe disability, respectively. In contrast, on the overall ADDS composite scale, 100% denoted normal, and 0% denoted severe disability.

Handwriting measurement assays:

All subjects provided writing samples using a pressure-sensitive pen on a digital tablet (MobileStudio Pro13, Wacom Co, Ltd, Japan). Subjects wrote at a self-selected pace and style (cursive or print) (Supplementary Videos 1 and 2). During the writing assessments, subjects copied the holo-alphabetic sentence “A large fawn jumped quickly over white zinc boxes” ten times in MovAlyzer kinematic software (Version 6, Neuroscript LLC, Temp, Az). The MovAlyzer software records the pen’s x, y, and z positions and the time function of the writing samples (Supplementary Figure 1A and 1B). The position parameters and time function are then transformed using a Fast Fourier algorithm to automatically calculate the kinematic features of writing in MovAlyzer (Supplementary Figure 1C–F). The kinematic writing features are also converted to numerical measures in the software which can be exported for statistical analysis (21,22). All the exported kinematic measures (Supplementary Table 1) were analyzed for a single stroke defined as the beginning to end of a sentence recording (23). The primary kinematic measures described in detail in this manuscript are peak accelerations which capture the dysfluent movements scored by clinicians and rated by patients and defined as the number of acceleration peaks in a single stroke, both up-going and down-going, and reported as counts. Normalized jerk is the change in acceleration during a single stroke. The jerk measure is normalized for stroke duration and size and, therefore, unitless. Straightness error represents the angular deviation in a single stroke and is calculated by taking the average distance to the minimum relative mean square line for each stroke. Straightness error is reported in cm (24,25). A visual representation of the writing kinematic features can be found on the company’s website (https://neuroscript.net/help/viewingtrials.html under “visualization of stroke features extracted by MovAlyzeR”).

Subjects also wrote the same holo-alphabetic sentence four times in OneNote software (Version 2016, Microsoft Office 365). Handwritten sentences were then converted to text using the automated writing-to-text conversion feature of the software. The writing-to-text feature applies a pattern recognition algorithm called optical character recognition (OCR) to convert the handwritten words to the closest approximate text (26). OCR algorithm is the basis of most artificial intelligence and machine learning softwares (27). The percentage of handwritten words correctly converted to text out of 36 words, regardless of punctuation, was measured by an analyst blinded to the group identity and labeled word legibility. The measure of word legibility is thus intended to capture the ability to communicate with writing that patients report as a disability. A visual representation of the writing-to-text feature is available on the company’s website (https://support.microsoft.com/en-us/office/write-notes-and-draw-in-onenote-for-windows-10-82d1189d-eb6d-4b07-9101-b50e13645c28).

All subjects performed the ten sentence copying assessments in MovAlyzer and four in OneNote softwares twice – an initial assessment (“pre-assessment”) and following a prolonged writing task (“post-assessment”). During the prolonged writing task, subjects performed a writing task in the MovAlyzer software consisting of copying a diverse holo-alphabetic paragraph for 20 minutes.

Statistical analysis of automated measures:

All automated measurements exported from MovAlyzer and OneNote softwares were tested for differences between WC and HV using two-tailed t-tests. All data were initially tested for normality using the Shapiro Wilk test. For data that passed normality, a student t-test was used for the two-tailed t-test. Mann-Whitney Rank Sum (MWRS) test was used for data that failed normality. Bonferroni post hoc analysis was used to correct for multiple comparisons across eight categories of writing measures with an adjusted p-value <0.00625 considered statistically significant. JMP Pro 16.0 software (SAS Co, SAS Institute Inc., Cary, NC) was used for these analyses. Continuous variables in the text are presented in the Tables as means (standard deviation [SD]) and in the graphs as box plots with medians (25th, 75th percentiles), minimum and maximum overlaid on the individual subject means.

Diagnostic performance analysis:

To characterize the diagnostic performance of automated measures, four logistic regression models were run modeling disease status – one for each automated outcome measure mean (peak acceleration, normalized jerk, straightness error, and word legibility). All models were controlled for age and gender. The model formula used for each of the four automated measures was:

$$\text{Logit}\text{(Disease)}={\beta }_0+{\beta }_{\mathit{\text{Age}}}\mathit{\text{Age}}+{\beta }_{\mathit{\text{Female}}}I\left(\mathit{\text{Female}}\right)+{\beta }_{\mathit{\text{automated measure}}}\mathit{\text{Automated measure}}$$

The diagnostic performance analysis was conducted using SAS 9.4 (SAS Institute Inc., Cary, NC). The associated receiver operating curve (ROC) was prepared and plotted for each of the four models. The area under the curve (AUC) was calculated for each of the four models.

Sample size calculation:

Two research designs were used to calculate sample size: 1) a parallel research design and 2) a single-arm crossover research design. A parallel research design compared differences between HV and WC using a two-tailed unpaired t-test. A single-arm crossover research design was used to detect the intra-subject difference between two time points (post-assessment and pre-assessment) in WC and calculate the standard deviation in a two-tailed paired t-test. In both research designs, an alpha of 0.05 and a power of 0.80 were used to estimate the sample size.

Test and retest reliability:

To evaluate for retest reliability, a smaller subset of subjects (n=8) from both groups performed three additional visits, at least one week apart. At each visit, subjects performed a handwriting measurement assay as detailed above. However, unlike the initial handwriting measurement assay, subjects only performed one writing assessment. To minimize the learning effect of copying the same sentence at repeat visits, subjects copied a different holo-alphabetic sentence at the three subsequent visits. At visits two, three, and four, subjects copied the following sentences respectively: “Pack my box with five dozen liquor jugs”; “the quick brown fox jumps over the lazy dog”; and “jinxed wizards pluck blue ivy from the big quilt.” Writing measures at visit one pre-assessment and repeat measures at visits two to four were used to calculate the intra-correlation class coefficient (ICC). Specifically, a crossed model type between subject and research visit was used to calculate the ICC for the writing measures of duration and peak accelerations. The ICC for word legibility could not be calculated because the quantitative approach of measuring total correct words from all four sentences resulted in insufficient technical replicates to measure ICC.

Correlation of automated measures with the subject and clinician-rated outcomes:

The automated measures of peak accelerations and word legibility were correlated with subject-rated outcomes that captured similar metrics. Specifically, word legibility was correlated with subject-rated writing legibility score on both BFM and ADDS scales. Peak accelerations was correlated with subject-rated BFM right arm dystonia (range 0-4) and total disability of BFM (range 0-30) and ADDS (0-100) scales. All correlational analyses were reported using Pearson’s correlational coefficient (R). To evaluate statistical significance, p-values were reported with all correlation coefficients. Bonferroni post hoc analysis was used to correct for multiple comparisons across five correlations with an adjusted p-value <0.01 considered statistically significant.

RESULTS

Clinical characteristics of WC

21 WC and 22 age-matched HV in equal proportion genders were enrolled in the study with no significant differences in mean age (WC: 59.4 years (SD 12.2) vs. HV: 59.8 years (SD 11.6), p=0.91, student t-test) or gender (WC: 12M/9F; HV 15M/7F, p=0.45, Chi-square test). WC showed a mean disease duration of 20.4 years (SD 12.3) with a range of four to 47 years. 28.6% of WC reported simple dystonia, while the remaining 71.4% endorsed dystonic symptoms with other fine motor tasks in addition to writing (Table 1). None of the healthy subjects reported any dystonic symptoms. WC enrolled in this study represent the disease population as previously reported, except for gender, which showed a greater male to female ratio and specifically four males to every three females in this study (4–6). Across the two examiner-rated scales, the mean WCRS part A total movement score was 6.2 (SD 3.8), WCRS part B writing speed was 0.8 (SD 0.7), and BFM right arm dystonia was 2.1 (SD 1.0). Across the two subject-rated scales, the mean BFM disability score was 2.8 (SD 1.1), and ADDS total score was 65% (SD 13.3).

Table 1:

Clinical Data of WC Subjects

Subject	Age (yrs)	Gender	WC Type	Dystonic tasks	Symptom Duration (yrs)	Examiner-rated			Subject-rated
						WCRS Part A	WCRS Part B	BFM R-arm Dystonia	BFM Disability	ADDS Disability
1	72	Male	Simple	-	4	1.5	1	1.5	4.0	56
2	63	Female	Simple	-	10	10.5	0	1.5	3.0	77
3	37	Male	Simple	-	11	8.5	1	1.5	2.0	64
4	73	Male	Simple	-	11	11.0	2	2.5	2.0	86
5	34	Female	Simple	-	20	3.0	0	0.5	1.0	90
6	59	Male	Simple	-	47	5.5	1	2.5	3.5	77
7	43	Male	Dystonic	typing, playing piano, using mouse	6	1.0	0	0.5	5.0	56
8	59	Male	Dystonic	typing	7	11.0	0	4.5	2.0	56
9	60	Female	Dystonic	typing, scissoring	35	8.0	2	2.5	3.0	NA
10	45	Female	Dystonic	typing	12	13.0	0	2.5	1.0	69
11	73	Male	Dystonic	typing	13	4.5	1	2.5	4.0	64
12	75	Female	Dystonic	gripping objects	14	8.0	0	2.0	3.0	64
13	72	female	Dystonic	typing, applying makeup	16	10.5	1	2.0	3.0	73
14	71	Male	Dystonic	gripping objects	17	5.5	2	3.0	2.0	51
15	65	Female	Dystonic	gripping objects	20	11.5	1	2.0	2.0	81
16	59	Male	Dystonic	typing, gripping objects	22	7.0	0	2.0	3.0	43
17	64	Female	Dystonic	applying makeup, using a spoon	24	3.0	0	1.5	2.0	69
18	49	Male	Dystonic	typing, gripping	30	11.5	1	3.0	5.0	51
19	53	Male	Dystonic	playing tennis	30	1.0	0	0.5	2.0	81
20	54	Female	Dystonic	typing, putting on jewelry, using utensils	36	3.5	1	3.0	3.0	43
21	67	Male	Dystonic	gripping	43	6.5	1	2.0	2.0	69

NA: missing data point, Dystonic tasks: all subjects reported dystonia with writing in addition to the tasks listed. WCRS: Writer’s Cramp Rating Scale, BFM: Burke-Fahn-Marsden, ADDS: Arms Dystonia and Disability Scale.

Four automated writing measures distinguish WC from HV

Twenty-two automated measures categorized into eight kinematic writing features were analyzed across 43 subjects using MovAlyzer software (Supplementary Table 1). Three automated measures showed differences between WC and HV: peak accelerations (Figure 1B; t(22)= −3.21; p=−0.004), straightness error (Figure 1C; U=125; p=0.006), and normalized jerk (Figure 1E; U= 118, p=0.004).

Figure 1: Three kinematic measures of writing distinguish WC from HV.

A) Peak writing dysfluency B) Angle of sentence writing C) Writing dysfluency in x and y-axes in WC and HV. Boxplots show median, inter-quartile ranges, maxima, and minima overlaid on participant mean with outliers (>2SD) shown as black dots outside boxplots. **denotes significance.

Two additional measures showed trends towards significant group differences including duration of writing (Figure 1A; t(24)= −2.59; N_HV=22, N_WC=21 p=0.016), normalized y-jerk (Figure 1D; U=132; p=0.01). Interestingly, additional automated measures previously reported in the literature (mean axial pen pressure, average velocity, stroke length, or CV of peak vertical velocity) did not show group-level differences (Supplementary Table 1). Next, word legibility was assessed using an automated handwriting recognition software (Figure 2A). Significantly fewer words were correctly recognized in writing samples from WC than HV (median U= 109; p=0.003) using this automated measure that directly reflects a clinically meaningful outcome of the ability to communicate via handwriting (Figure 2B).

Figure 2: Automated word legibility can distinguish WC from HV.

A) Automated writing-to-text conversion software was used to convert the handwritten sentences from HV and WC (left column) to text (right column). Incorrect words for each text writing sample are shaded in gray. B) WC demonstrate lower word legibility compared to HV.

Four automated measures show high diagnostic performance

Next, the sensitivity and specificity of these automated measures to distinguish WC from HV were evaluated using a logistic regression model controlled for age and gender (Figure 3). The ROC curves show that all four measures have high sensitivity and specificity to differentiate between the two groups, albeit normalized jerk (AUC 0.79; CI: 0.66, 0.93) and word legibility (AUC: 0.77; CI: 0.62, 0.93) show higher diagnostic performance compared to the measures of straightness error (AUC: 0.72; CI: 0.56, 0.88), and peak accelerations (AUC: 0.72; CI: 0.54, 0.89).

Figure 3: Four automated writing measures show high sensitivity and specificity in predicting WC status.

For each automated measure, a logistic regression test was run to model WC status, and the diagnostic performance of each measure was plotted (ROC curve). ROC curves above the diagonal gray line show high sensitivity and specificity to predict WC status. An aggregate diagnostic performance (AUC) measure is also provided with 95% confidence intervals in the parenthesis.

Two automated measures show a realistic sample size to detect group differences

The sample size needed for each of the four automated measures to detect a clinical effect was next evaluated in two standard clinical research designs. A single-arm crossover research design in WC would allow the study of the effect of a research manipulation or clinical intervention in the same disease subjects. Although the definition of clinically meaningful effect size can vary widely due to study aims, we used a medium Cohen’s d-effect size of 0.3 to evaluate the sample size in this research design. The estimated sample sizes of WC needed to see a clinical effect for each measure were: word legibility- 32; peak accelerations- 58; normalized jerk- 84; and straightness error- 300. As a comparison, the estimated sample for examiner-rated measures of WCRS part A movement scores- 43; part B writing speed- 142 and BFM right arm dystonia severity- 6.

A parallel research design in WC and HV would compare the effect of a research manipulation or clinical intervention between the two groups. The estimated sample size of each group needed to see a clinical effect for each measure was: word legibility- 23; peak accelerations-18; normalized jerk- 25; and straightness error- 43. The estimated sample size of each group for each examiner-rated measure was: WCRS part A movement score- 15; part B writing speed- 51 and BFM right arm dystonia- 9. Overall, the sample size calculations across the two research designs showed a wide range, with smaller sample sizes estimated for automated measures of word legibility and peak accelerations and larger sample sizes estimated for measures of normalized jerk and straightness error.

One automated measure shows high retest reliability

One automated measure with a smaller sample size (peak accelerations) was evaluated for test and retest reliability for subjects from both groups. Across four writing assessments, each performed at least one week apart, the measure of peak accelerations showed a high ICC of 0.80. Although the measure duration of writing did not meet the threshold for group differences between HV and WC, it showed a high ICC within subjects of 0.96. Overall, an analysis of the kinematic measures of peak accelerations showed high test and retest reliability and, therefore, can be reliably used across multiple research visits.

Two automated measures correlate with subject-rated outcomes

To evaluate the clinical meaning of automated measures, the two automated measures with the smaller sample sizes (word legibility and peak accelerations) were correlated with subject-rated outcome measures that captured similar metrics. Specifically, the automated word legibility measure inversely correlated with subject-rated writing legibility in BFM (Figure 4A, R=−0.51, p=0.0005) and ADDS scales (Figure 4B, R= −0.45, p=0.003). Peak accelerations correlated with subject-rated BFM right arm dystonia score (R= 0.48, p=0.001) and to a lesser extent subject-rated disability scores as captured by BFM scale (Figure 4C, R= 0.38, p=0.01) but not ADDS scale (R=−0.35 p=0.02). Overall, the automated measures of word legibility and peak accelerations captured similar phenomenology as traditionally rated by subjects and clinicians, thus allowing us to understand the clinical impact of research findings using these metrics.

Figure 4: Two automated measures correlate with subject reported outcomes.

Word legibility Measure inversely correlates with subject reported A) BFM and B) ADDS writing legibility scores. Higher numbers on the x-axis and lower numbers on the y-axis correspond with decreased legibility. Peak accelerations correlate with subject reported C) BFM right arm dystonia and D) BFM disability scores. Pearson’s correlation coefficient (R) and p-value are reported. n=21 WC and 22 HV.

DISCUSSION

In summary, this study identified two automated writing measures (peak accelerations and word legibility) that (1) differentiate WC from HV, (2) show high sensitivity and specificity, (3) achieve a realistic sample size for a rare disease, (4) correlate with the traditional subject and clinician-rated outcomes, and (5) the kinematic measures of duration and peak accelerations also demonstrate high test and retest reliability for use in a clinical research study.

A key finding of this study is that measures of writing kinematics and specifically dysfluency can be used as a quantitative measure for WC disability in clinical research studies. Writing kinematics has also been studied in Parkinson’s disease (PD), with a better understanding of the link between kinematic measures and clinical disease. Specifically, Teuling and colleagues showed that PD subjects show greater variability in the practiced skill of writing than HV due to reduced coordination and motor control of finger, wrist, and arm movements. Teulings and colleagues then developed kinematic writing measures to capture the writing variability of PD subjects in the MovAlyzer software also used in this study. They showed that the normalized jerk measure could capture PD subjects’ writing variability (17). WC subjects in this study could also be differentiated from HV using the measure of normalized jerk. However, in this study, the measure peak accelerations captured a similar phenomenology but showed a better signal-to-noise ratio in WC. Measuring writing dysfluency using an automated measure such as peak accelerations is clinically important because it captures the phenomenology of dystonia reported by the traditional clinician and subject-rated outcome measures. Specifically, WCRS total movement score includes measures of dystonic posture and latency of dystonia, while BFM right dystonia and disability scale captures subject-rated disability. Therefore, peak accelerations capture an important dystonia feature traditionally captured in both clinician-rated and subject-rated outcome measures.

To our knowledge, our study also provides the first evaluation of a consumer-accessible writing recognition software as an outcome measure for word legibility in WC. Previous studies have used circle drawing to evaluate the accuracy of writing movements and have shown group-level differences (10,11,28). This study’s automated measure of word legibility captures group-wise differences in writing accuracy previously reported (10,11,28) but is a more proximal measure of the clinically meaningful outcome of being able to communicate with writing, as demonstrated by its inverse correlation to two subject-rated writing disability scores. Furthermore, of the two automated measures identified here, word legibility using this consumer software (OneNote, Microsoft) requires the smallest estimated sample size to detect a moderate effect in the two clinical research designs. Future studies using word legibility should increase the number of technical replicates to five to ten duplicate sentences per writing assessment to calculate the test and retest reliability score.

While our findings reinforce some prior observations, our study did not identify group differences in other primary measures previously reported, such as mean stroke frequency, mean axial pressure, and CV of peak velocity (10–12,16). However, there are methodological differences between this study and prior studies. Examples of methodological differences are the sensitivity of the digital writing tablet and software, the parameters of the writing assay (the writing sample used, number of writing trials, duration permitted for each trial), and the approach to data analysis (e.g., mean vs. CV of the sentence velocity). Overall, this study supports two automated writing measures that show high sensitivity and specificity to distinguish WC from HV and show a high signal-to-noise ratio to achieve a realistic sample size in this rare disease.

A link between kinematic writing measures and basal ganglia dysfunction in PD subjects was previously demonstrated (17). Teuling and colleagues used both actual writing samples and simulated writing samples from PD subjects to demonstrate a link between abnormal writing patterns in PD and basal ganglia dysfunction. Specifically, reduced coordination and motor control of PD subjects in their study were shown to result from decreased dynamic modulation of the central movement generator in the basal ganglia, which progressively declines with the loss of dopamine signal (17). Their analysis of the kinematic writing measures showed that measures of dysfluency (specifically normalized jerk) could specifically link the clinical manifestation of PD to basal ganglia dysfunction. In the future, it would be interesting to determine if the kinematic measure of writing dysfluency in WC (peak accelerations) captures not only the clinical features of the disease but also links with elements of the disease pathology such as basal ganglia dysfunction. Future clinical research studies should compare functional brain changes with kinematic writing measures in WC to understand the link between functional brain activity and writing behavior changes in dystonia. Understanding the link between functional brain activity and kinematic writing measures may help us dissociate primary functional brain abnormalities from secondary brain changes in dystonia. An understanding of writing kinematic measures in WC may thus be a gateway to understanding disease mechanisms in all focal dystonias.

In conclusion, this study identifies two automated writing measures: a novel word legibility measure and a kinematic measure of peak accelerations in WC dystonia research studies. At present, therapeutic opportunities for individuals affected by WC are limited, and affected individuals suffer from both an impaired ability to communicate via writing and pain. A critical step towards improving the therapeutic landscape for WC and broadly other forms of focal dystonias are validated outcome measures suitable for testing small populations and relevant to clinically meaningful outcomes. The use of validated automated outcome measures such as the ones presented in this research study can allow for a more effective research design to advance disease mechanisms and therapies in this disabling and rare disorder.