Integrating remote testing and machine learning to identify markers of cerebellar ataxia at home

Abstract

Background

In-person assessments face accessibility, scalability, and geographic diversity challenges, especially for rare diseases. Additionally, Cerebellar Ataxia (CA) non-motor symptoms(NMS) are often overlooked. We aimed to address these gaps by leveraging the Internet and machine-learning.

Methods:

In a bi-center study, we assessed 100 participants: 30 CA, 45 neurotypically healthy(NH), and 25 Parkinson’s disease(PD), recruited from 57 geographical locations across two countries. We evaluated multiple domains—cognition, anxiety, depression, social support, and personality—using accessible online tools. We applied leave-one-out cross-validation and feature importance analysis to examine the machine-learning model’s ability to distinguish between groups and identify the most sensitive and specific CA predictors.

Results

Machine-learning models trained on these remote non-motor features alone, yield AUCs of 0.74/0.76(CA vs. NH) and 0.78/0.79 (CA vs. PD) using leave-one-out cross-validation, demonstrating classification power exceeding 20%.

Conclusion

These findings highlight the value of integrating digital-health technologies and machine-learning models for CA NMS evaluation, potentially serving as scalable digital-markers.

Plain language summary

Cerebellar ataxia (CA) is a neurological condition that affects movement, but people with CA often have problems beyond movement, such as changes in thinking and mood. People with CA find it difficult to travel to clinics, making assessments challenging. We tested whether these non-movement symptoms could be measured remotely using online platforms. We compared people with CA with healthy controls and people with Parkinson’s disease. We found that people with CA showed a distinct pattern of non-movement symptoms, especially in cognitive abilities and personality traits. Computer programs trained solely on these online measures achieved medium accuracy in identifying people with CA. These findings suggest that online assessments could improve access to care and research for people with CA, especially those who live far from medical centers.

Ponger et al. combine remote non-motor assessments with machine learning to identify cerebellar ataxia across geographically diverse populations. Their findings support the potential of remotely acquired non-motor digital markers as scalable tools for ataxia detection and monitoring.

Introduction

Cerebellar ataxia (CA) is a rare neurodegenerative condition affecting less than 0.01% of the population¹. CA comprises a heterogeneous group of disorders, including hereditary, sporadic, and acquired etiologies, which differ in clinical presentation and progression. Current in-person assessment methods for CA present several challenges². In this study, we aimed to address challenges related to accessibility, scalability, geographical diversity, and disease-specific sensitivity.

Assessment accessibility is a major challenge for people with CA. Many patients experience motor and mobility impairments that make travel to clinics or labs difficult³. Gait difficulties and poor coordination can further reduce the capacity of individuals with CA to attend or fully participate in testing. Cognitive and affective symptoms may compound these barriers, as these can interfere with traveling and assessment procedures.

Scalability remains a significant limitation, as most assessments require administration by experts, such as neurologists or neuropsychologists, with restricted availability in remote or rural areas, thus further limiting repeated or large-scale assessments⁴. Recruitment challenges also frequently lead to small samples, often fewer than 30 participants, drawn from single institutions, raising concerns about sampling bias and the generalizability of research findings⁵.

Importantly, for CA patients who live in rural areas or under-resourced regions, expert access is limited. Major metropolitan cities benefit from greater access to medical centers, which can lead to geographic and socioeconomic disparities in diagnosis and monitoring. Accordingly, many previous studies have employed geographically limited samples, often recruiting participants from a single medical center. This over-reliance on metropolitan populations limits the generalizability of findings across geographically and culturally diverse populations. Collectively, these challenges constrain both clinical care and research efforts, underscoring the need for more accessible and scalable methods globally.

Additionally, CA has traditionally been defined and diagnosed based solely on motor symptoms. Consequently, in-person assessments for CA typically focus mainly on motor measures. However, growing evidence shows that CA has Non-Motor Symptoms (NMS), such as cognitive deficits^4,6–8, neuropsychiatric conditions, and personality changes^6,9,10. While NMS affect quality of life and disease management, they are often overlooked and under-recognized in clinical assessments¹¹. Current data suggest that nearly 95% of individuals with CA experience at least one neuropsychiatric symptom⁶. For example, individuals diagnosed with CA can develop cerebellar cognitive affective syndrome¹² (CCAS). Studies consistently show lower scores on the Montreal Cognitive Assessment (MoCA) compared to NH, indicating mild cognitive impairment (MCI) in this population⁴. CA is also sometimes associated with higher prevalence of anxiety, depression¹³, and personality trait changes¹⁴, such as neuroticism¹⁵ and impulsivity^16,17.

However, it is not clear which NMS are sensitive and specific markers to distinguish CA from both neurotypically healthy (NH) individuals and other movement disorders, such as Parkinson’s disease (PD). Identification of the most discriminative non-motor markers is critical not only for understanding the clinical phenotype of CA, but also for the design of potential screening tools that prioritize the most informative NMS domains^6,9,10,18,19.

To address these challenges, we leverage the Internet. This online approach could offer both time- and cost-effective solutions, as the tests can be easily administered in the participant’s home environment. For our study, we selected platforms that are free, widely used, and easily accessible (e.g., Google Forms). This approach reduces barriers for administrators by simplifying setup and minimizing the need for technical support. This approach could provide a method of evaluating NMS in CA, reducing barriers to participation and enabling larger, more representative samples. It could also give clinicians and researchers scalable tools for early detection and tailoring interventions in both research and clinical settings. Remote evaluation of NMS is more accessible for patients with early to mid-stage CA, who are still able to use digital technology. Thus, we focused on this patient population.

It remains unknown which non-motor measures are sensitive and specific to CA when assessed remotely. Therefore, we first tested these measures using online tools to identify those that differentiate CA from other groups. We assessed multiple domains: cognition, anxiety, depression, social support, and personality traits using simple online tools. We examined three groups: individuals with CA, neurotypically healthy (NH), and Parkinson’s disease (PD). This dual-comparison design enabled us to assess both the sensitivity and specificity of online non-motor measures for identifying CA. The inclusion of PD participants, representing another neurodegenerative motor disorder with recognized NMS, allowed us to evaluate whether the observed patterns were CA-specific rather than reflecting general neurodegenerative disease processes.

Second, using machine-learning (ML), we evaluated whether non-motor measures alone could support accurate identification of CA. We developed an ML classifier to evaluate whether remotely collected non-motor measures can accurately distinguish individuals with CA from the two comparison groups (NH and PD) based solely on the non-motor measures. Given the typical small sample size in this rare disease population, we used leave-one-out cross-validation (LOOCV) to assess our ML classifier’s predictive power in distinguishing between all group pairs, based solely on the online collected NMS.

Third, we examined which features had the most predictive power in CA classification. Typically, in studies of CA, this group is assessed in only one or two domains. However, by utilizing our online data collection approach, we were able to administer an extended assessment battery encompassing multiple domains, including cognition, anxiety, depression, social support, and personality traits. Among these, to identify the most predictive NMS, we used the LOOCV with permutation feature importance analysis, examining which predictors were most sensitive and specific to CA.

Our findings demonstrate that remotely collected non-motor measures differentiate individuals with CA from both NH individuals and patients with PD. Machine-learning models trained on these online measures achieve moderate classification performance, supporting the potential of scalable digital assessments to aid remote identification of CA and advance digital markers for clinical research and future care.

Methods

Participants

A total of 100 individuals were tested: 30 with a previously established diagnosis of CA, 45 NH, and 25 PD. Participants were recruited through the Ataxia clinic in the Department of Neurology, Tel Aviv Sourasky Medical Center (TLVMC), and the Center for Accessible Neuropsychology (CAN) database. The participants’ demographic and medical information are presented in Table 1.

Table 1.

Demographic and medical information of all groups. Mean [SE] (Range)

Group	N	Age	Years of education	Females (%)	Motor	Disease Duration
Cerebellar ataxia	30	54.7 [11.9] (30–73)	15.7 [3.1] (12–25)	76.6	13.6 [6.4] (3–25.5)	4.6 [4.7] (1–19)
Parkinson’s Disease	25	61.1 [8.51] (45–77)	15.8 [2.3] (12–22)	68	18.6 [10.78] (4.8–56.2)	5.9 [5.37] (1–18)
Neurotypical Healthy	45	58.5 [12.5] (34–87)	15.6 [2.6] (9–21)	71.1	NA	NA

Inclusion criteria included age above 18, ability to understand and provide informed consent, and access to an internet connection with a video conferencing device. Exclusion criteria included additional neurological disorders to CA or PD, psychiatric conditions, learning disabilities, severe visual or auditory impairments, and deep brain stimulation amongst PD participants. Tel Aviv University (#0005713–6) and TLVMC (TLVMC 0761–20) Institutional Review Board Committees approved the protocol, and all participants provided informed consent.

The CA group was comprised of 24 participants with diverse etiologies (1 ARSACS, 1 SCA1, 3 SCA2, 16 SCA3, 1 SCA42, 1 SCA6, 1 SCA8) and six with CA of unknown etiology. PD participants were tested in the ON state. All participants were tested without any change to their current medication regimen. The CA group had a mean disease duration of 4.6 years and a mean SARA score of 13.6, indicating mild to moderate impairment^20,21. The PD group had a mean disease duration of 5.9 years and MDS-UPDRS III mean score of 18.6, which also typically reflects mild to moderate impairment²².

Importantly, participants were recruited from 57 distinct geographical locations across the USA and Israel (CA = 18, NH = 18, PD = 21). See Fig. 1.

Fig. 1. Geographical marker map of the participants from 57 distinct geographical locations across the USA and Israel.

Map data ©2025 Google, Imagery ©2025 TerraMetrics. CA = Cerebellar ataxia; NH = Neurotypical healthy; PD = Parkinson’s disease.

Procedure

Neurological and neuropsychological assessment

We followed the online neuropsychological testing protocol, called iPONT^13,19,20. Individuals were invited by email to participate in an online, live interview with an experimenter. After providing informed consent, the participant completed a demographic and medical questionnaire. For the CA and PD participants, the experimenter obtained the participant’s medical history, collecting information about age at diagnosis, medication, primary symptoms, genetic subtype, diet, other neurological or psychiatric conditions, and more relevant information. The experimenter then administered the Montreal Cognitive Assessment test (MoCA)²³ as a brief evaluation of cognitive status. For NH participants, this session ended with the completion of the MoCA.

Next, we assessed motor severity, administering the Scale for Assessment and Rating of Ataxia (SARA)²¹ to the participants with CA, and the motor section of the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS III)²² to the participants with PD. Similar to all assessments, SARA and MDS-UPDRS-III were administered remotely via structured video-based examinations, following protocols from published work on remote motor assessment in these populations^{20,24,25,26,27}. This session took 40–60 min to complete.

Subsequently, we administered digital questionnaires assessing anxiety, depression, social support, and personality traits in a counterbalanced and random order^13,20,27. Participants completed all questionnaires at their convenience on their personal computers within two weeks of the initial assessment.

Anxiety and depression were assessed using the Patient-Reported Outcomes Measurement Information System (PROMIS)⁵. Anxiety was assessed utilizing the PROMIS anxiety scale, and depression levels were measured with the PROMIS depression scale. The PROMIS is scored with a t-score metric (mean is 50, and the standard deviation is 10). Social support was evaluated using the Interpersonal Support Evaluation List (ISEL-12), which consisted of 12 items²⁸. The ISEL-12 also included three sub-scales: Appraisal, Tangible, and Belonging. The PROMIS and ISEL-12 were validated in CA and PD²⁰. Personality traits were assessed using the NEO International Personality Item Pool (IPIP-NEO-120)²⁹. This instrument evaluates the Big Five personality dimensions: neuroticism, extraversion, openness, agreeableness, and conscientiousness.

Importantly, for our study, we selected online platforms that are free, widely used, and easily accessible. This approach reduces barriers for administrators by lowering costs, simplifying setup, and minimizing the need for technical support. Such advantages improve scalability and facilitate integration into future research and clinical settings. The PROMIS and ISEL-12 questionnaires were administered via Google Forms, while the IPIP-NEO-120 was administered via Qualtrics.

Data analysis

Analysis was conducted using R software or JASP software. The sample size was determined based on previous comparable work in CA and PD patients.^{PD: 30–34; CA: 35–39}. To calculate the required sample size, we conducted a power analysis (alpha = 0.05; power = 0.99) using effect sizes derived from five studies for each group (PD: d = 1.152, large effect; CA: d = 1.265, large effect).^{PD: 30–34; CA: 35–39}. These analyses suggested a minimal sample size of 14 participants for each group (PD = 13.39; CA = 11.38). As such, the sample sizes of our groups (≥25) had sufficient power to detect group differences.

To test the group differences between the CA and NH groups, we performed a two-sample independent t-test for each domain. As this was an exploratory study, we did not correct for multiple comparisons. To compare PD and CA, given that the ages of the PD and CA groups were significantly different (typically, PD patients are older), we used analysis of covariance (ANCOVA) with age as the covariate variable. We did not find an effect of age in the comparison between CA and PD in any domain (p > 0.05). Consequently, the Results section presents the group effect. Also, we conducted a chi-square test for goodness of fit between the percentage of MCI, clinical anxiety, and clinical depression between the groups (MoCA score <26 for MCI, PROMIS t-score > 54 for anxiety and depression).

Machine-learning models for the identification of CA remotely

Our ML models were implemented in Python using scikit-learn (Version 1.6.1). We applied logistic regression with L2 (ridge) regularization and Random Forest classifiers, balancing simplicity, interpretability, and robustness to small sample sizes⁴⁰. Logistic regression estimates the probability of group membership by modeling the log-odds of the outcome as a linear function of the input features. The L2 penalty constrains the size of the model coefficients, helping to reduce overfitting by discouraging overly complex solutions and improving generalization to unseen data. Random Forest is an ensemble learning method that constructs multiple decision trees and aggregates their predictions, enabling the capture of non-linear relationships and interactions between features⁴¹. The additional classifier was implemented to assess consistent results across models.

All preprocessing steps, including z-score normalization, were performed within each training fold during LOOCV and then applied to the held-out test sample, thereby preventing data leakage. The feature set included: MoCA score, anxiety and depression t-scores, the five traits of the BIG-5, and ISEL-12’s three sub-scores.

We employed the leave-one-out cross-validation (LOOCV) algorithm to thoroughly evaluate our ML classifier’s ability to differentiate between CA and NH, CA and PD, or PD and NH individuals. This method is particularly suitable for small datasets, providing an estimate of classifier performance and helping prevent overfitting in limited sample sizes⁴². Classification performance was evaluated using standard metrics: area under the receiver operating characteristic curve (AUC-ROC), accuracy, precision, and recall. Uncertainty was quantified using bootstrap resampling with 1000 iterations, and 95% confidence intervals of the AUC were computed using the percentile method⁴³.

Feature importance analysis of the machine-learning models

Feature importance was assessed separately for the CA vs. NH and CA vs. PD comparisons, allowing us to explore which predictors were most sensitive and specific to CA. Feature importance was examined using permutation importance analysis derived from the models fit to the full dataset. In this approach, the values of each feature were randomly permuted while holding the trained model fixed, and the resulting decrease in classification performance (AUC) was used as an estimate of that feature’s contribution to the model. Features associated with the largest performance decrease were interpreted as the most important contributors to classification. This allowed us to identify the non-motor measures that contributed most strongly to the ML model performance.

Results

See Fig. 2 for a summary of the results in a Radar Plot, combining all domains for a comprehensive profile of each group. See full statistical results in Tables 2–4.

Fig. 2. Radar plot of the scores of all three groups in all domains, in standardized Z scores.

CA = Cerebellar ataxia; NH = Neurotypical healthy; PD = Parkinson’s disease.

Table 2.

Full statistical information for comparisons between the CA group to NH group in all domains, using two-sided t-tests. Mean [SD]

Domain	CA	NH	t (73)	P	Effect size
Cognition	25.2 [2.7]	27 [1.7]	−3.667	<0.001*	0.864
Depression	52.6 [8.2]	50.2 [7.9]	1.291	0.201	0.304
Anxiety	52.6 [8.8]	49.8 [7.9]	1.406	0.164	0.331
Neuroticism	65.8 [15.8]	61.4 [11.4]	1.413	0.162	0.333
Extraversion	69.9 [16.1]	79.2 [11.6]	−2.912	0.005*	0.686
Conscientiousness	96.2 [10.05]	93.1 [11.7]	1.167	0.247	0.275
Openness	78.8 [12.5]	81.2 [13.6]	-0.755	0.453	0.178
Agreeableness	98.4 [8.1]	91.1 [12.8]	2.774	0.007*	0.654
Social support-Total	25.5 [6.9]	27.5 [5.5]	−1.393	0.168	0.239
Social support-Appraisal	8.5 [3.04]	9.7 [2.1]	−2.010	0.048*	0.474
Social support-Tangible	9.1 [2.3]	9.6 [2.5]	−0.992	0.324	0.234
Social support-Belonging	7.9 [2.9]	8.3 [2.1]	−0.823	0.413	0.194

* = p < 0.05; CA = Cerebellar ataxia; NH = Neurotypical healthy.

Table 4.

Full statistical information for comparisons between the PD group to NH group in all domains, using two-sided t-tests. Mean [SD]

Domain	NH	PD	t (68)	P	Effect size
Cognition	27 [1.7]	26.2 [2.26]	−1.715	0.091	−0.428
Depression	50.2 [7.9]	52.3 [7]	1.126	0.264	0.281
Anxiety	49.8 [7.9]	54.4 [5.87]	2.509	0.014	0.626
Neuroticism	61.4 [11.4]	62.7 [11.63]	0.462	0.645	0.115
Extraversion	79.2 [11.6]	76.5 [8.29]	−1.035	0.304	−0.258
Conscientiousness	93.1 [11.7]	86.4 [15.08]	−2.085	0.041	–0.520
Openness	81.2 [13.6]	76.8 [9.75]	–1.425	0.159	−0.355
Agreeableness	91.1 [12.8]	81.24 [18.1]	−264	0.010	−0.659
Social support-Total	27.5 [5.5]	27 [6.74]	−0.370	0.712	−0.092
Social support-Appraisal	9.7 [2.1]	9.3 [2.47]	−0.734	0.465	−0.183
Social support-Tangible	9.6 [2.5]	8.8 [2.5]	−1.259	0.212	−0.314
Social support-Belonging	8.3 [2.1]	8.8 [2.67]	0.735	0.465	0.183

* = p <.05; NH = Neurotypical healthy; PD = Parkinson’s disease.

Cognition

As expected⁴, we found a significant difference between the CA and NH groups in the MoCA score (Table 2; Fig. 1-S), such that the CA group received a lower score than NH. We did not find a significant difference between the CA and PD groups (Table 3) and between the NH and PD groups (Table 4). Moreover, we evaluated the percentage of clinical MCI (<26) for each group. We found a significant group difference in the proportion of participants classified with MCI (χ²(2, N = 100) = 12.881, p =.002), such that the CA group showed a higher percentage of MCI than the NH group.

Table 3.

Full statistical information for comparisons between the CA group to PD group in all domains, using two-sided t-tests. Mean [SD]

Domain	CA	PD	F (1,52)	P	Effect size
Cognition	25.2 [2.7]	26.2 [2.26]	1.795	0.186	0.033
Depression	52.6 [8.2]	52.3 [7]	0.094	0.760	0.002
Anxiety	52.6 [8.8]	54.4 [5.87]	1.198	0.279	0.022
Neuroticism	65.8 [15.8]	62.7 [11.63]	0.703	0.405	0.013
Extraversion	69.9 [16.1]	76.5 [8.29]	4.121	0.047 *	0.073
Conscientiousness	96.2 [10.05]	86.4 [15.08]	9.359	0.004 *	0.152
Openness	78.8 [12.5]	76.8 [9.75]	0.176	0.677	0.003
Agreeableness	98.4 [8.1]	81.24 [18.1]	21.012	<0.001 *	0.582
Social support-Total	25.5 [6.9]	27 [6.74]	1.077	0.304	0.02
Social support-Appraisal	8.5 [3.04]	9.3 [2.47]	1.202	0.278	0.023
Social support-Tangible	9.1 [2.3]	8.8 [2.5]	0.004	0.948	0.000
Social support-Belonging	7.9 [2.9]	8.8 [2.67]	2.221	0.142	0.04

* = p < 0.05; CA = Cerebellar ataxia; PD = Parkinson’s disease

Depression and anxiety

We did not find a difference in anxiety score between CA and NH groups (Table 2 and Fig. 2-S) and between the CA and PD groups (Table 3). We found a significant difference in the anxiety score between NH and PD, where the PD group received a higher score (Table 4). In terms of depression, we found no difference in the depression score between all pairs of groups (Tables 2–4). We did not find a significant group difference in the proportion of participants classified with clinical anxiety (χ²(2, N = 100) = 3.932, p = 140) or depression (χ²(2, N = 100) = 0.857, p = 0.652).

Personality Traits

The CA group received a significantly higher score than the NH group in the Agreeableness trait and a lower score in Extraversion (Table 2 and Fig. 4-S). We found no significant difference in the other traits between the two groups (Table 2). When comparing CA and PD, we found an effect of group on Extraversion, Agreeableness and Conscientiousness (Table 3), such that CA was higher than PD in Agreeableness and Conscientiousness, but lower on Extraversion. There was no effect of group on the other traits (Table 3). When comparing the PD and NH groups, we found a significant difference in Agreeableness and Conscientiousness (Table 4) where PD received a lower score than the NH group. We did not find differences in other traits (Table 4).

Social support

We found that the CA group scored significantly lower than the NH group in the social support appraisal sub-score only (Table 2 and Fig. 3-S). No group differences were found in belonging or tangible sub-scores, or the ISEL-12 total score (Table 2). No group differences between CA and PD were found in any social support score (Table 3). We did not find significant differences between PD and NH in any social support score (Table 4).

Machine-learning models for the identification of CA remotely

Logistic regression

When distinguishing CA from NH, the logistic regression model with L2 regularization yielded an AUC of 0.74 and an overall accuracy of 70.5% (AUC 95% CI: 0.56–0.88, precision = 70%, recall = 71%). These results indicate moderate discriminatory performance, providing a 24% improvement over a random model (AUC = 0.50). Notably, this discrimination power was achieved using only remote non-motor measures, emphasizing the utility of such features in differentiating between CA and healthy individuals. For PD vs. NH classification, the ML model achieved an AUC of 0.65 and an accuracy of 70.4% (AUC 95% CI: 0.44–0.84, precision = 69%, recall = 70%), reflecting modest predictive power with a 15% gain in classification power.

In the CA vs. PD classification, the ML model demonstrated modest classification performance, with an AUC of 0.78 and an overall accuracy of 68.4% (AUC 95% CI: 0.57–0.92, precision = 68%, recall = 78%). This corresponds to a 28% increase in classification power over a non-informative model. These results highlight the discriminative capacity of digital non-motor markers in separating between two neurodegenerative clinical groups. See Fig. 3 for the full results.

Fig. 3. Receiver operating characteristic (ROC) curve and area under the curve (AUC) for the machine-learning logistic regression classifiers (CA vs. NH, PD vs. NH, and PD vs. CA) using leave-one-out cross validation.

CA = Cerebellar ataxia, NH = Neurotypical healthy, PD = Parkinson’s disease.

Random forest

Using a random forest classifier, classification of CA vs. NH yielded an AUC of 0.76 and an overall accuracy of 70.5% (AUC 95% CI: 0.58–0.88; precision = 65%; recall = 59%), indicating moderate discriminative ability. This represents a 26% improvement over a non-informative baseline (AUC = 0.50). For the PD vs. NH comparison, the ML model achieved an AUC of 0.65 and an accuracy of 68% (AUC 95% CI: 0.46–0.85, precision = 56%, recall = 36%), suggesting low-to-moderate classification performance and an approximate 15% increase over chance-level prediction.

In differentiating CA from PD, the random forest model achieved an AUC of 0.79 and an overall accuracy of 75.4% (AUC 95% CI: 0.55–0.93; precision = 78%; recall = 60%), representing a 29% improvement over a non-informative model. See Fig. 4 for the complete results.

Fig. 4. Receiver operating characteristic (ROC) curve and area under the curve (AUC) for the machine-learning random forest classifiers (CA vs. NH, PD vs. NH, and PD vs. CA) using leave-one-out cross validation.

CA = Cerebellar ataxia, NH = Neurotypical healthy, PD = Parkinson’s disease.

Feature importance analysis of the machine-learning models

Logistic regression

Analysis of feature importance revealed distinct patterns in the contribution of non-motor digital markers to group classification: CA vs. NH, PD vs. NH, and PD vs. CA (Fig. 5). For the CA vs. NH classification, MoCA, agreeableness, and extraversion scores emerged as the strongest predictors (mean decrease in AUC = 0.12, 0.11, and 0.09, respectively), indicating that cognition and personality traits play a central role in distinguishing individuals with CA from healthy controls in remote assessments.

Fig. 5. Feature importance analysis of the machine-learning logistic regression models.

Distinct patterns of non-motor markers contributing to group classification are shown for CA vs. NH, PD vs. NH, and PD vs. CA comparisons. CA = Cerebellar ataxia, NH = Neurotypical healthy, PD = Parkinson’s disease.

In the PD vs. NH classification, anxiety, social support belonging, and agreeableness emerged as the strongest predictors (mean decrease in AUC = 0.17, 0.12, and 0.11, respectively). In the PD vs. CA classification, agreeableness, anxiety, and MoCA were the strongest predictors (mean decrease in AUC = 0.18, 0.04, and 0.04, respectively).

Random forest

Feature importance analysis of the Random Forest classifier revealed distinct patterns in the contribution of non-motor digital markers to group classification: CA vs. NH, PD vs. NH, and PD vs. CA (Fig. 6).

Fig. 6. Feature importance analysis of the machine-learning random forest models.

Distinct patterns of non-motor markers contributing to group classification are shown for CA vs. NH, PD vs. NH, and PD vs. CA comparisons. CA = Cerebellar ataxia, NH = Neurotypical healthy, PD = Parkinson’s disease.

For the CA vs. NH classification, extraversion, agreeableness, and MoCA scores emerged as the most influential predictors (mean decrease in AUC = 0.04, 0.003, and 0.001, respectively). These findings align with the logistic regression results, highlighting the central role of both cognition and personality traits in distinguishing individuals with CA from healthy controls in remote assessments. In the PD vs. NH comparison, extraversion, agreeableness, and anxiety were identified as the leading predictors (mean decrease in AUC = 0.009, 0.004, and 0.005, respectively), suggesting the role of personality and affective features in distinguishing PD from healthy individuals. For the CA vs. PD classification, agreeableness, extraversion, and conscientiousness were the most important features (mean decrease in AUC = 0.03, 0.006, and 0.0001, respectively), indicating that personality-related measures contribute meaningfully to differentiating between these two neurodegenerative conditions.

Discussion

In this study, we evaluated which non-motor measures are sensitive and specific to individuals with early to mid-stage CA when assessed remotely. By leveraging accessible and freely available online platforms, we aimed to reduce logistical barriers and enable remote participation. Using this method, and in contrast to previous studies that measure one or two domains¹³, we assessed multiple non-motor domains, including cognition, anxiety, depression, personality traits, and social support. We examined three groups: individuals with CA, neurotypically healthy (NH), and Parkinson’s disease (PD).

Second, using two kinds of machine-learning (ML) classifiers, we evaluated whether non-motor measures alone could support accurate identification of CA. We developed an ML classifier to evaluate whether remotely collected non-motor measures can accurately distinguish individuals with CA from the two comparison groups (NH and PD) based solely on their non-motor measures. Given the small sample size in this rare disease population, we used leave-one-out cross-validation (LOOCV) to assess our ML classifier’s predictive power in distinguishing between all group pairs, based solely on the online collected NMS. Third, we examined which features had the most predictive power in CA classification, potentially serving as digital markers.

We found that the CA group has a distinct combination of NMS that significantly differentiates them from both NH individuals and patients with other neurodegenerative conditions (i.e., PD) when assessed remotely. CA participants scored significantly lower on cognition compared to NH participants, reinforcing the understanding that cerebellar dysfunction affects cognition early in the disease course⁴⁴. Also, the CA group reported reduced social support in the “appraisal” domain. In terms of personality, the CA group were less extraverted and more agreeable than both NH and PD groups. Although previous studies have reported a higher frequency of depression⁴⁵ and anxiety⁶ in CA, our findings did not show increased prevalence or significant group-level differences when assessed remotely.

First, mean levels may appear similar even when the frequency of clinically significant symptoms is higher. Second, our CA sample consisted primarily of individuals with mild-to-moderate disease severity, which might have contributed to lower levels or frequency of emotional difficulties. Third, participants were assessed remotely in their home environment, which could have reduced symptoms of depression and anxiety compared to in-person assessments. Considering these factors, future studies should directly compare in-person and online administration methods to determine whether the observed results reflect sample characteristics (such as disease type or severity) or differences in administration methods.

The ML models, trained on remotely collected non-motor measures, demonstrated moderate discriminatory power in distinguishing CA from both NH and PD groups, in both types of classifiers (AUC = 0.74/76 and 0.78/79, respectively). The non-motor measures added more than 20% in classification power over a naïve model, indicating their value in detecting disease presence and differentiating between other neurodegenerative conditions. These results emphasize that non-motor measures can support identification of disease presence and offer clinically relevant diagnostic information even in the absence of motor information. Importantly, our ML models' performance relied solely on digital assessments that are brief and scalable for wide deployment, including among rare disease populations. Future work should systematically compare a broader range of ML models in order to identify the most robust and generalizable models for disease classification across different datasets and clinical settings.

Among all measures, cognition, extraversion, and agreeableness were the most predictive features for distinguishing CA from NH controls, which implies that cognitive impairment and personality alterations could play an important role in the non-motor phenotype of CA. In contrast, the CA vs. PD comparison highlighted agreeableness, anxiety, and MoCA, suggesting that personality traits, emotional status, and cognition might better capture divergence between these two neurodegenerative conditions, possibly reflecting different cerebellar and extra-cerebellar network involvement^6,17.

Five limitations should be acknowledged. First, the CA group was not stratified by genetic subtype⁴⁶. Given the phenotypic heterogeneity across subtypes (e.g., SCA1, SCA2, SCA3), this may have limited the detection of subtype-specific non-motor profiles. CA comprises a heterogeneous group of disorders with hereditary, sporadic, and acquired etiologies that differ in clinical presentation and progression. Accordingly, our findings should be interpreted within this heterogeneous clinical framework. Stratifying by disease type in future studies could help clarify these distinctions and support more targeted, personalized approaches to care. However, this remains challenging, as CA is a rare condition and sample sizes are typically small (n < 20). Future research may benefit from larger, multi-center samples stratified by CA subtype.

Second, the use of self-report measures (e.g., for depression) may have been influenced by individual subjective bias or response tendencies⁴⁷. Although suitable for accessible screening, the absence of clinician-rated or performance-based measures limits the ability to have more objective measures or capture more dimensions of the disease. However, expert administration is costly and not accessible. Third, in the search for sensitive digital markers of CA, future work should examine a broader range of non-motor symptoms, including fatigue and impulsivity^6,12,16,48. Identifying the most discriminative NMS could help develop a sensitive, remote assessment panel for CA. The cerebellar cognitive affective syndrome (CCAS) scale¹² is a validated bedside instrument specific to cerebellar populations. Although not yet adapted for remote use, several domains assessed in our study overlap with CCAS components (e.g., cognition). Extending disease-specific assessment into a scalable home-based format is an important future goal.

Fourth, participants completed the assessments independently in naturalistic settings, which may have introduced variability in testing conditions and additional noise (e.g., distractions, device type, internet connection, or environmental context). Finally, although participants were instructed to complete assessments independently, remote environments may introduce potential noise and third-party influence, which cannot be fully controlled. This limitation reflects a broader trade-off between ecological validity and experimental control that is inherent to remote testing paradigms. Future studies should directly compare in-person and online administration methods to evaluate their equivalence and identify potential differences.

Our online approach provides a standardized method for administering multi-domain assessments to people who speak English or Hebrew. This approach requires only an internet connection, a personal digital device, and a webcam, and does not require a caregiver’s presence while still ensuring participant safety. Additionally, the tests can be administered by non-experts, reducing reliance on in-person evaluations by experts and providing a time- and cost-effective method in two countries: the USA and Israel.

This online approach may support longitudinal monitoring, integration into clinical trials as digital assessments, and the delivery of remote cognitive rehabilitation and psychosocial interventions. While these findings highlight the potential of digital-health approaches for remote assessment, further large-scale and externally validated studies are required before establishing broad clinical utility.

The current results demonstrated the potential of integrating accessible digital-health technologies and ML algorithms to evaluate NMS in CA remotely. This approach could offer an accessible and scalable method for classifying CA’s NMS, and, ultimately, other neurodegenerative conditions. In the long run, we hope that remote neuropsychological tests will enable more accessible assessments across diverse communities, languages, and diseases. The current findings can pave the path for accessibility, scalability, and geographical diversity in CA assessments.

Ethics statement

This study was approved by the Tel Aviv University Institutional Review Board committee (#0005713-6) and was conducted in accordance with the Declaration of Helsinki (TLVMC 0761-20). Informed consent was obtained from all participants. Clinical trial number: not applicable.

Author contributions

W.S. conceived the study. D.M. and S.B. collected the data. Y.D.P. analyzed the data. W.S., Y.D.P., and P.P. interpreted the results. H.L. and W.S. drafted the manuscript. All authors reviewed, edited, and approved the final version of the manuscript.

Peer review

Peer review information

Communications Medicine thanks Myunggi Yi who co-reviewed with Hadi Malekroodi and Roberto Rodríguez-Labrada for their contribution to the peer review of this work.

Funding

This study did not receive external funding.

Data availability

All data generated by this study are included in this paper and its supporting files. Source Data for figures may be found in Supplementary Data 1.

Code availability

The code used for the ML analysis is available at https://github.com/yaeldepicciotto/Integrating-remote-testing-and-machine-learning-to-identify-markers-of-cerebellar-ataxia-at-home.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s43856-026-01794-1.