2.1 Literature Search and Review

We first searched the PubMed/MEDLINE database for relevant literature published in the English language, between the initial descriptions of COVID-19 (December 31, 2019) and January 25, 2021, using the search terms “COVID-19” OR “SARS-CoV-2” OR “Coronavirus Disease 2019” OR “2019 n-CoV” OR “2019 Novel Coronavirus” AND (“movement disorders” OR “myoclonus” OR “ataxia” OR “parkinsonism” OR “Parkinson’s disease” OR “chorea” OR “dystonia” OR “myoclonus” OR “catatonia” OR “tremor”)” as search terms (n = 332). We added filters for Case Reports, Classical Articles, Clinical Studies, Clinical Trial, Comparative Studies, Original Journal Articles, Letters, Observational Studies, Human Species and identified 296 records for screening (titles/abstracts). We conducted a prospection in the reference list of selected papers and other sources, identifying additional 19 items. After excluding 115 records with abstracts and titles irrelevant to this study’s scope, we screened 200 records for eligibility. We evaluated 81 full-text papers for eligibility after withdrawing 119 articles for the following reasons: they were not written in English or were focused on social or healthcare impact of COVID-19 in patients already diagnosed with a movement disorder, dealt with aggravation of pre-existing movement disorders, telemedicine management strategies, or were listed as review/hypothesis/opinion articles (Figure 1). A total of 44 studies were included for qualitative and quantitative synthesis as representative of neurological descriptions of new-onset movement disorders in the context of SARS-CoV-2 infection.

2.2 Data extraction, synthesis, and analysis

Two authors (PB and TG) collected data independently, using a predefined protocol. One of the five authors (PB) checked the extracted data, and discrepancies were resolved by consensus between the two data collectors. Variables of interest included counts for single or combined movement disorders (myoclonus, dystonia, ataxia, parkinsonism, chorea, catatonia, action/postural tremors, oculomotor disturbances), associated clinical features (encephalopathy, stroke), summaries of ancillary exam results (FDG-PET, MRI/CT, ENMG, CSF, and EEG). The attempted treatment options (symptomatic treatment or immunotherapy, such as intravenous immunoglobulin – IVIg, steroids) were also counted. Clinical characteristics were quantified as case counts and ratios of case counts to total cases. Symptoms not initially reported in individual studies were considered absent (zero) rather than missing (NA).

Descriptive statistical analysis was performed using Wizard v2.0.4 (Evan Miller, 2020, https://www.wizardmac.com) and R v4.0.4 (https://www.R-project.org/). A world map including the reported cases was generated using Datawrapper (https://app.datawrapper.de). The “brms” R package was used to obtain the parameter values’ Bayesian posterior distribution [5]. “Brms” uses the logit-link to convert the linear predictor zi (zero-inflated) to a probability. The logit-link accepts values between 0 and 1 and returns values on the real line. As a result, it enables the conversion of probabilities to linear predictors. Samples were drawn using No-U-Turn-Sampler (NUTS) sampling, the default algorithm used by the Stan program [6]. Rhat was used as a potential scale reduction factor on split chains (at convergence, Rhat = 1). We fit the model using four Hamiltonian Monte-Carlo (HMC) chains. We used Jeffrey’s beta (0.5, 0.5) as the prior argument for zi rather than the default “flat” beta (1, 1) used by “brms”. The intercept parameter prior was normal (0, 10). The Bayesian plot was created using the “bayesplot” R package (https://mc-stan.org/bayesplot/) [7].

Using counts as the measure of movement disorders frequency (main variables extracted from the articles), we performed an unsupervised grouping procedure (clustering) to separate and observe co-occurrence between the symptoms in the reported articles. The “NbClust” R package was used to determine the optimal number of clusters [8]. It provides 30 indices for that task and proposes the best clustering scheme by varying all combinations of several clusters, distance measures, and clustering methods (Supplementary Figure 1). We then employed a random forest (RF) clustering strategy to avoid variable transformation (e.g., categorical features). By forming two unique groups (as determined by the NbClust’s method), RF clustering separated the original data from the virtual clusters. The similarity between two data instances was assessed using the proportion of trees that shared a single leaf node. The PAM (partition around medoids) technique [9] was used to assign each object to the closest medoid (object with the smallest dissimilarity) as a step to create the clusters.

3.1 Studies characteristics

Forty-four initial papers were ultimately chosen from a total of 81 full-text articles that were evaluated for eligibility, including a total of 93 specific cases of movement disorders caused by or acutely linked to SARS-CoV-2 infection or its pharmacological treatment attempts. Table 1 summarizes these cases and publications.

The cases came generally from regions where COVID-19 epidemics were on course. A proportion of 57.0% of the cases were reported in Europe (n = 53), 26.9% in the Americas (n = 25), 15% in Asia (n = 14), 1% in Oceania (n = 1). Most reported cases came from France (n = 17, 18.3%), United States (n = 17, 18.3%), Spain (n = 16, 17.2%), followed by Brazil (n = 8, 8.6%), Iran (n = 8, 8.6%) Germany (n = 8, 8.6%), Italy (n = 7, 7.5%), United Kingdom (n = 3, 3.2%) and China (n = 3, 3.2%). Japan, Netherlands, Israel, New Zealand, Philippines, and Belgium reported 1 case each (1.1% each), as represented in Figure 2.

The major findings are depicted in Figure 3 as Bayesian binomial posterior proportions. Table 1 summarizes the clinical and ancillary exam data from each original study to serve as a simple reference guide for physicians and researchers.

3.2 Clinical characteristics

Of the 93 movement disorder cases identified in the literature search [3, 4, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51], myoclonus was present in 63.4% (n = 59), ataxia in 38.7% (n = 36), action/postural tremor in 10.8% (n = 10), rigid-akinetic syndrome (parkinsonism) in 5.38% (n = 5), oculomotor abnormalities (opsoclonus, ocular flutter) in 20.4% (n = 19), catatonia in 2.1% (n = 2), dystonia in 1.1% (n = 1), chorea in 1.1% (n = 1), functional (psychogenic) movement disorders in 3.2% (n = 3) of the reported COVID-19 cases with any movement disorder (Figure 2). Encephalopathy was a frequent association (n = 37, 39.8%), and stroke was detected in a minority of cases (n = 5, 5.4%).

As previously stated, the most prevalent movement disorder was myoclonus [4, 11, 13, 14, 15, 16, 17, 18, 27, 28, 30, 31, 32, 33, 34, 39, 41, 43, 47, 48, 50, 51, 52], which often occurred in conjunction with encephalopathy/delirium. More than a third of the sample had ataxia [16, 17, 18, 19, 20, 21, 22, 23, 27, 29, 32, 33, 34, 37, 38, 39, 40, 43, 46, 50, 51], which often co-occurred with myoclonus. Oculomotor symptoms, such as saccadic oscillations (opsoclonus, ocular flutter), have also been documented and were often associated with ataxia [13, 14, 16, 18, 19, 27, 29, 41, 43, 47, 50, 51]. Among the cases presenting with ataxia, a few were associated with acute polyradiculopathy/neuropathy (sensory ataxia, in the context of Miller-Fisher syndrome) [19, 20, 37]. In contrast, others presented cerebellar ataxia in a cerebellitis context [23] or unremarkable MRI and CSF examinations [16, 17, 18, 21, 25, 27].

Not surprisingly, confusional or encephalopathic states were highly prevalent in patients with action and stimulus-sensitive myoclonus. Myoclonus is a frequent involuntary movement in critically ill patients triggered or aggravated by several different drugs or by metabolic or hypoxia-related brain dysfunction [53]. Encephalopathy and delirium are very prevalent in critical care [54], and the reviewed sample has a significant proportion of hospitalized and critical patients. Three cases of new-onset rigid-akinetic parkinsonism were described in detail [10, 11, 26], one of which presented with encephalopathy and generalized myoclonus [11], two of which were preceded only by a mild respiratory syndrome [10, 26].

Some de novo movement disorders were presumably drug-induced: two serotonin syndrome cases possibly related to lopinavir/ritonavir combined with drugs with serotonergic properties, and three tardive syndrome cases (described as rigid-akinetic syndromes combined to orobuccal stereotypies) [13]. Due to incomplete information, it was not feasible to establish a causal link between myoclonus and drug administration in critically ill mechanically ventilated patients, in whom sedative agents, especially opioids, or antipsychotic medications, may have contributed to myoclonus.

3.3 Neuroimaging (MRI, CT, FDG-PET, and functional dopaminergic imaging)

Neuroimaging (Brain MRI or head CT) findings were reported in 82.8% (n = 77) of the 93 cases. About two-thirds (n = 56, 72.7%) of these cases demonstrated normal or non-related findings (chronic findings, such as white matter small vessel disorder). Five patients had cerebral microbleeds (6.5%). Stroke was disclosed in 4 cases (5.2%). Post-contrast focal enhancement or edema was seen in 10 cases (12.9%). Some of these observations are described in greater detail below: Gadolinium (Gd) enhancement of the oculomotor nerve (in a Miller-Fisher syndrome case, 1/77, 1.3%), cerebellar edema with Gd enhancement (1/77, 1.3%), cortical or brainstem hypoxic-ischemic lesions (2/77, 2.6%), pachymeningeal enhancement (1/77, 1.3%), acute leukoencephalitis (3/77, 3.9%), splenium DWI hyperintensity (1/77, 1.3%), splenium microbleeds (1/77, 1.3%), fornix edema (1/77, 1.3%), orbitofrontal hyperintensity (1/77, 1.3%), middle cerebral artery stroke (2, 2.6%), and multiple focal enhancing lesions in putamen and cerebellum (1, 1.3%).

FDG-PET scan was performed in five patients, as described below. It showed normal glucose metabolism in a case with parkinsonism [10]. Two of these five scanned patients had myoclonus and ataxia [34, 43], one had ataxia and other cerebellar signs [43], and one had orofacial myoclonus [47]. FDG-PET revealed hypermetabolic cerebellum in all four cases: two also had hypermetabolism in the basal ganglia, while the other three had cortical hypometabolism.

The three parkinsonism cases had confirmed nigrostriatal denervation by functional dopaminergic imaging, using 18F-DOPA PET, DAT-Scan, or TRODAT-1 SPECT.

3.4 Cerebrospinal fluid analysis

For 48.4% (n = 45) of the examined cases, cerebrospinal fluid (CSF) analysis was available. About 73.3 % (n = 33) of the patients had a normal cell count, while 26.7% (n = 12) had mild-to-moderate pleocytosis. We observed that 51% (n = 23) of patients had normal CSF protein levels, while 49.9% (n = 22) had slightly elevated protein levels.

Antineuronal paraneoplastic antibodies (Abs) were not detected in any of the cases when a standardized Abs panel was tested (n = 9). Antibodies against Purkinje cells, striatal, and hippocampal neurons were observed in one patient with subacute cerebellar syndrome and myoclonus utilizing tissue immunostaining [34]. Franke et al. used indirect immunofluorescence to screen antibodies against novel CNS epitopes in unfixed mouse brain sections. They found staining associated with vessel endothelium, astrocytic proteins, and neuropil of basal ganglia, hippocampus, or olfactory bulb [47].

SARS-CoV-2 RT-qPCR positivity in CSF samples appears to be uncommon: it was detected in only one case of cerebellitis [23].

3.5 Neurophysiological techniques (EEG and EMG)

EEG findings were available in 29% (n = 27) of the cases. Of these, 59.3% (n = 16) had background slowing, while 33.3% (n = 9) had normal results. In one case, lateralized periodic discharges were observed in association with diaphragmatic myoclonus [31], while in another, frontal spikes were observed [21]. Two studies (three cases) used neurophysiological measures to characterize the possible CNS topographic origin of myoclonus [15, 31], using a combination of EEG and EMG or jerk-locked registers. Two of the three cases had cortical/subcortical origin, and one, a subcortical source.

3.6 Treatment attempts

In 39.8% of the cases (n = 37), the information about treatment attempts was incomplete or missing. There was immunotherapy with oral/intravenous steroids (n = 21, 37.5%). In 11 cases, the corticosteroid therapy targeted the movement disorder; in 5 cases, it targeted COVID-19-related inflammatory reaction and, in 4 cases, the target was not clear. Immunoglobulin (IVIg) was used in 32.14% (n = 18) of the reported cases with available treatment information (n = 56). In all of these cases, IVIg therapy was targeting the movement disorder. Plasma exchange was tried in one case. Immunotherapy was used primarily in inpatient settings, but cases ranged from mild to critically ill in mechanical ventilation.

Antiepileptic drugs and benzodiazepines such as clonazepam (n = 17), levetiracetam (n = 16), valproic acid (n = 10), and lorazepam (n = 5) seemed to be favored for symptomatic care, mostly for the myoclonus presentation. Additionally, dexmedetomidine (n = 5), ketamine (n = 1), primidone (n = 1), and midazolam (n = 1) were used in treatment attempts.

3.7 Bayesian regression using zero-inflated Poisson distribution

In this analysis, we generated a regression using Bayesian zero-inflated Poisson (ZIP), a model for counts-based datasets with an excess of zero-valued data in the dependent variable (assuming a constant zero-inflation probability across observations). A value of “zero” indicates that nothing occurred, which may be because the events’ rate is low or because the process that causes events failed to start.

Within the data extracted from the sample of COVID-19 associated with new-onset movement disorders, myoclonus was associated with normal neuroimaging results (head CT or brain MRI) and encephalopathy (left side of Figure 4), where the posterior densities of the beta distributions (obtained from the Bayesian zero-inflated Poisson regression) do not exceed zero (central line). When ataxia was used as the outcome variable (right side of Figure 4), no relation was observed (all posterior beta densities crossed zero) with the dependent variables.

3.8 Distinct syndromes associated with myoclonus

Cluster analysis was utilized to infer movement disorders syndromic grouping from the reviewed articles. According to a seven indices ensembling method (NbClust), we found two as the optimal number of clusters. The PAM method then evaluated dissimilarity among objects. By assembling two unique groups, the random forest (RF) clustering approach [53] identified 36 articles with a higher proportion of myoclonus reports associated with encephalopathy and eight articles with a higher proportion of myoclonus connected with ataxia or oculomotor disturbances.

We interpreted these findings as suggestive of at least two myoclonus-associated syndromes that might occur in COVID-19: (1) a syndrome that associates myoclonus and encephalopathic state (which, we speculate, may relate to with broader brain dysfunction and neuronal hyperexcitability due to clinical complications in critical care or pharmacological interactions); and (2) a post-viral syndrome with an autoimmune origin that links oculomotor disturbances, ataxia, and myoclonus, resembling an opsoclonus-myoclonus-ataxia syndrome.