xHD-Vox, an Automated Speech Model for Estimating Motor and Cognitive Scores in Huntington Disease: Development and Longitudinal Validation

Introduction

Huntington disease (HD) is a rare autosomal neurodegenerative disorder caused by an expanded cytosine-adenine-guanine (CAG) repeat (>35) in the huntingtin gene [1]. The disease manifests through a variety of motor, cognitive, and psychiatric symptoms, requiring lifelong multidisciplinary care. While no cure is currently available despite extensive research on disease modifiers, symptomatic treatments are effective in managing disability for many years. Therefore, continuous and personalized monitoring of disease progression is essential to optimize patient care. However, access to specialized expertise remains limited, highlighting the need for scalable, automated approaches to support clinical decision-making.

The composite Unified Huntington disease Rating Scale (cUHDRS) is a robust metric for tracking disease progression, integrating motor, cognitive, and functional scores from the UHDRS [2]—the worldwide reference scale for HD clinical evaluation. Its assessment is typically conducted once a year during face-to-face examinations by trained neurologists and neuropsychologists with restricted availability [3]. This calls for cost-effective and remote-friendly solutions to enable continuous disease monitoring.

Speech analysis has emerged as a promising, noninvasive, and practical tool for tracking HD progression [4,5]. While other biomarkers, such as striatal atrophy on magnetic resonance imaging, cerebrospinal fluid neurofilament levels, and cognitive batteries, have been validated to monitor disease progression, speech-based assessments are well-suited for remote, home-based implementation, reducing the logistical burden and costs associated with in-clinic visits. In our previous study [5], we demonstrated that speech features could predict cUHDRS scores and their cognitive and motor components more accurately than the combination of demographic and genetic data [5]. These features were extracted from a simple counting task, in which participants counted forward and backward from 1 to 20 in less than 45 seconds. However, the need for manual annotations by speech pathologists still limited the model’s scalability and implementation as a telemonitoring solution [6,7].

To address these limitations, we developed xHD-Vox (x for eXplainable, HD for Huntington disease, and Vox for voice), an interpretable and automated model that predicts cUHDRS and its motor and cognitive components from brief speech recordings in HD gene carriers. This study comprises the following two main parts: (1) xHD-Vox development, where we ensured interpretability by using a standard linear regression model and selecting a subset of relevant speech features identified in our previous work [5]. Manual annotations were replaced with features extracted using Whisper, an open-source automatic speech recognition (ASR) system developed by OpenAI, known for its high transcription accuracy and multilingual support [8]. (2) xHD-Vox evaluation on an independent longitudinal dataset to evaluate its ability to track disease progression over time.

Methods

Participants

Participants were enrolled from three prospective longitudinal studies: BIO-HD (NCT01412125), REPAIR-HD (NCT03119246), and MIG-HD (NCT00190450) before transplants.

We selected data from HD gene carriers with a CAG repeat length of greater than or equal to 40 to ensure the complete penetrance of HD [9]. We excluded visits with missing speech recordings or clinical scores required to compute cUHDRS. This resulted in the selection of 181 participants, with a total of 341 visits (Figure 1).

For the evaluation of our model on an independent longitudinal test set, we selected participants who had 3 consecutive annual visits, which resulted in a cohort of 24 participants (72 visits). Assessing performance over a 2-year follow-up period ensured a sufficiently long observation window to capture disease progression.

The remaining 157 participants, comprising 269 visits (mean 1.7, SD 1.1), were used for model training.

Among these, 145 recordings from 90 participants were annotated by speech pathologists and used to develop the automated model.

‎

Clinical Evaluation

Participants were assessed by annually certified examiners with the Unified Huntington Disease Rating Scale (UHDRS). For this study, we used the cUHDRS and its four components to evaluate the disease progression: Total Motor Score (TMS), the Symbol Digit Modalities Test (SDMT), the Stroop Word (SW), and the Total Functional Capacity (TFC). The cUHDRS was calculated using the following formula [2]:

${\displaystyle {\displaystyle \begin{array}{rl}\mathrm{c}\mathrm{U}\mathrm{H}\mathrm{D}\mathrm{R}\mathrm{S}& =\frac{\mathrm{T}\mathrm{F}\mathrm{C}-10.4}{1.9}-\frac{\mathrm{T}\mathrm{M}\mathrm{S}-29.7}{14.9}\\ & +\frac{\mathrm{S}\mathrm{D}\mathrm{M}\mathrm{T}-28.4}{11.3}+\frac{\mathrm{S}\mathrm{W}-66.1}{20.1}+10\end{array}}}$

The CAG-age-product (CAP) score was calculated using the standardized formula calibrated to reach 100 at the expected age of onset [10]:

${\displaystyle {\displaystyle \mathrm{C}\mathrm{A}\mathrm{P}=\mathrm{a}\mathrm{g}\mathrm{e}\times \frac{\mathrm{C}\mathrm{A}\mathrm{G}-30}{6.49}}}$

Standardized Lightweight Speech Task

Speech samples were recorded in a clinical research environment at Henri Mondor Hospital through 2 brief controlled tasks. Each participant was asked by a neuropsychologist to (1) count aloud numbers from 1 to 20 (forward counting), then (2) to count backward from 20 to 1 while holding their hands up and closing their eyes (backward counting). The backward task increases cognitive load by inhibiting the forward counting process [5]. Recordings were made using an external microphone (Zoom H4n Pro; 44 kHz recording, resampled at 16 kHz) for the BIO-HD and REPAIR-HD cohorts and were extracted from the audio track of video recordings for MIG-HD. The microphone was placed on the neuropsychologist’s desk, facing the patient, who was seated approximately 1‐2 meters away.

For the annotated samples, speech recordings were transcribed by speech therapists at the word level using the Praat (developed by Paul Boersma, David Weenink, and Anastasia Shchupak) and Seshat (developed by Hadrien Titeux, Rachid Riad, Xuan-Nga Cao, Nicolas Hamilakis, Kris Madden, Alejandrina Cristia, Anne-Catherine Bachoud-Lévi, and Emmanuel Dupoux) platforms [11]. Mispronunciations (using the Speech Assessment Methods Phonetic Alphabet), paraphasias, phoneme perseverations, abnormal breathing, vocal noises, and filled pauses (“euh” and ”um”) were identified.

Analysis Implementation

This study followed the TRIPOD+AI (Transparent Report of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) reporting guideline. We conducted the analysis in Python using the scikit-learn library [12]. Features were normalized with RobustScaler to enhance feature importance analysis and improve the model’s convergence. For cross-validation, we generated 50 folds, with each fold containing 80% of participants in the training set and the remaining 20% in the test set. Each participant was assigned exclusively to either the training or test set within each fold.

Automated Model Development

Overview

In this first analysis, we automated the model developed by Riad et al [5] for predicting cUHDRS and its components using a parsimonious approach to reduce model complexity while preserving predictive accuracy. This approach involved the following three steps: (1) ranking the most predictive features of Riad et al [5] model, (2) identifying the smallest subset of features that achieved at least the same performance as the full feature set by progressively adding features in order of importance, and (3) automating the annotation process for the selected features using Whisper.

Ranking the Most Predictive Features of our Previous Model

Features Investigated

We applied the automatic pipeline for feature extraction based on speech pathologists’ annotations previously developed by Riad et al [5]. This process resulted in the extraction of 60 speech features (30 features for each of the 2 tasks), including articulatory and phonatory deficiencies, rhythm and temporal statistics, filled pauses, vocalization additions, and sequence-related metrics (eg, errors in number order and perseveration errors) (Table S1 in Multimedia Appendix 1 provides the detailed list of features extracted). We excluded the feature phones per second due to their high correlation (r=0.99) with pronounced numbers per second. Demographic features, including CAP score, age, and CAG repeat length, were added, leading to a total of 63 features being investigated.

Features Importance Ranking

Riad et al [5] model is a 63-feature ElasticNet regression, a standard linear regression with L1 and L2 regularization to prevent overfitting and mitigate multicollinearity. The study demonstrated that using only features from the backward task yields performance comparable to that achieved using features from both forward and backward tasks (ablation study). Performance was assessed using mean absolute error (MAE) across 50 cross-validation folds, which quantifies the average absolute difference between observed and predicted clinical scores.

Given similar results in our replication (Tables S2 and S3 in Multimedia Appendix 1), we retained only features from the backward counting task for further selection.

To identify the most relevant features from this task, we calculated the average absolute value of the ElasticNet regression coefficients across 50 cross-validation training folds. Features were then ranked by their relative importance for each clinical score prediction. Finally, we computed an average rank for each feature across cUHDRS, TMS, and cognitive scores. TFC was excluded from this average ranking because the model showed low explained variance (R²) for this score, indicating poor predictive performance.

Identifying the Minimal Predictive Features Set

To identify the minimal set of features that achieved performance at least equivalent to that of the initial model [5], features were progressively added in order of importance, as determined in step (1), in a standard linear regression model. We stopped adding features once the model’s performance matched or exceeded that of the initial model. Performance was evaluated using the mean MAE across 50 cross-validation folds. This method identified a minimal subset; however, it may not be globally optimal, as other combinations of features could potentially provide better predictive performance.

Automating the Selected Features’ Annotation Process Using Whisper

The two selected features in step (2) were “numbers pronounced per second” and “SD of the duration of individual numbers.” Numbers pronounced per second were computed as the total number of numbers spoken (including perseverations and sequence errors) divided by the task duration. Task duration was defined as the time interval from the start of the first number pronounced to the end of the last number pronounced. The SD of duration was calculated across the duration of individual spoken numbers.

To automate the extraction of these 2 features, it was necessary to identify the sequence of numbers produced by each participant within recordings that could include hesitations, fillers, or unrelated speech (eg, questions addressed to the neuropsychologist). This required accurate transcription of the recordings as well as precise text-audio alignment to compute timing-related features.

We used Whisper, a pretrained ASR model developed by OpenAI [8], which provides both transcription and timestamp alignment, thereby facilitating automated extraction of timing features. Whisper was trained on 680,000 hours of labeled audio, including multilingual datasets, and has demonstrated a low word error rate in French (word error rate=5.3 on FLEURS and 10.8 on Common Voice [version 15]). Its robustness to noisy recordings and coverage of 96 languages made it well suited for this task.

Numbers pronounced per second and the SD of the duration of individual numbers were compared using Pearson and Spearman correlations between the values computed from Whisper’s annotations and those derived from speech-language pathologist annotations.

Finally, we validated our automated model by comparing the MAE of predictions based on features extracted from speech pathologists’ annotations with those obtained using Whisper’s alignment.

Model Evaluation Using Longitudinal Data

Overview

We evaluated our automated model on longitudinal data using a leave-one-out setup to simulate its real-world performance. The model was calibrated on the training set before being tested on unknown participants (test set).

All longitudinal assessment results are reported on the test set to ensure evaluation on an unseen cohort and avoid inflated performance estimates. Given the test set selection approach, we verified a posteriori that (1) the training set contained enough participants for proper model calibration, (2) the training/testing split followed an 80/20 ratio in terms of visits, and (3) the test and training sets were balanced in terms of disease severity (refer to the Participants subheading in the Results section for details).

Model Performance on the Test Set

The model was trained using linear regression, and its performance was evaluated on the test set using the MAE, R², and intraclass correlation coefficient (ICC). The R² metric quantifies the proportion of variance explained by the model, whereas the ICC measures the agreement between predicted and observed clinical scores. To highlight the potential contribution of speech features, we compared the performance of our model with that of a baseline demographic model (based on CAP score and CAG repeat length).

Functional Classification

Given the low performance of the regression model for TFC, we explored functional decline using a logistic regression model trained to classify stage 3 vs stages 0‐2 of the Huntington Disease Integrated Staging System (HD-ISS). The model included two predictors: CAP score and speech rate (measured as numbers pronounced per second). Selection of these predictors was supported by cross-validation results (Figure S3 and methods in Multimedia Appendix 1). Model performance was evaluated using Receiver Operating Characteristic–Area Under the Curve (ROC AUC), accuracy, and recall.

Longitudinal Change

To test the hypothesis of a longitudinal decline in clinical scores, both as measured by clinicians and as predicted by the model on the test set, we performed 2-way repeated-measures ANOVAs separately for each score. The 2 within-subject factors were visit time (baseline, year 1, and year 2) and type of measure (clinician-assessed vs model-predicted). Statistical significance was set at α=.05, and Bonferroni corrected for the 5 tested scores.

For each ANOVA, we examined the main effects of time and type of measure, as well as their interaction. When a significant effect of time was observed, Tukey post hoc comparisons were conducted to characterize differences between visits.

To quantify the magnitude of longitudinal change, we computed 1-year and 2-year changes in each score for both clinician-assessed and model-predicted measures. One-year change was defined as the difference between the score at year 1 and baseline, whereas 2-year change was defined as the difference between the score at year 2 and baseline. Group-level agreement between clinical and predicted changes was assessed by comparing mean changes and their 95% CIs.

To evaluate model performance at the individual level, we examined the relationship between predicted and clinically measured 1-year changes using scatterplots, Pearson correlation coefficients, and Spearman correlation coefficients. This analysis was restricted to 1-year changes to maximize the number of observations (48 paired changes), thereby increasing statistical power. Sensitivity to outliers was assessed and reported in Multimedia Appendix 1.

Ethical Considerations

All participants signed an informed consent. Ethical approval was obtained from the institutional review board of Henri Mondor Hospital for BIO-HD and MIG-HD and of Saint-Louis Hospital for the French part of REPAIR-HD. The study complied with the Declaration of Helsinki, current Good Clinical Practice guidelines, and local laws and regulations.

Results

Participants

The full database (N=181 participants, 341 visits) included individuals at early stages of the disease, with low to moderate functional impairment (mean TFC 11.2, SD 2.3). Detailed demographic and clinical characteristics are provided in Table 1.

This database was split into training and testing sets (refer to Figure 1 for the flow of subset construction). The test set comprised 21% of the total visits and was balanced in terms of disease severity compared with the training set: 50% of participants were at stage 3 in both sets, and 19% were at stage 0‐1 (vs 32% in the training set). Clinical scores (cUHDRS, TFC, SDMT, SW, and TMS) were also close between the 2 subsets.

Our sample size analysis (Figure S2 in Multimedia Appendix 1) indicated that 100 participants were sufficient to calibrate the model with 95% accuracy; this condition was met with 157 participants in the training set.

The speech pathologist-annotated database used for model development showed a disease severity profile similar to that of the full database, with a slightly lower mean TFC (mean 10.8, SD 2.3).

A PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram of data selection is provided in Figure S1 in Multimedia Appendix 1. Only 1.2% of clinical data were missing.

Automated Model Development

Ranking the Most Predictive Features

Table S3 in Multimedia Appendix 1 shows that the R² for the TFC score was low (0.13) compared with that of the other scores, which were closer to 0.5. Although we continued to report results for TFC, it was excluded from the feature selection process because of its poor predictive performance.

The four most predictive features, based on their average rank, were the CAP score, CAG repeat length, and 2 rhythm-related speech features: numbers pronounced per second and the SD of the duration of pronounced numbers. Table 2 shows the top 15 most predictive features from the backward speech task for predicting clinical scores.

Identifying the Minimal Predictive Features Set

Table 3 shows that using the top 4 ranked features identified in step 1.1 (Table 2) resulted in MAE values that were equal to or lower than those of the initial 63-feature model (column 6 vs column 1). Using only the 2 top-ranked features (CAP score and numbers pronounced per second) was already sufficient for cUHDRS and SDMT to achieve lower MAE compared with the 63-feature model (column 1 vs column 4). Adding CAG repeat length further reduced the MAE for TMS (column 5 vs column 4), whereas incorporating the SD of the duration of pronounced numbers leads to an additional decrease in MAE for SW prediction (column 6 vs column 5).

The final 4-feature model outperformed a demographic baseline model based on CAP score and CAG repeat length (column 6 vs column 2; paired t test with Bonferroni correction, P<.001 for all scores). R² values, reported in Table S4 in Multimedia Appendix 1, complement the MAE analysis: for example, the final model achieved a mean R² of 0.59 for cUHDRS (vs 0.43 for the baseline) and 0.57 for SW (vs 0.37).

The demographic baseline includes only the demographic features present in the final model to evaluate the incremental value of speech features. Age was not included as a sensitivity analysis showed that adding age to the demographic features did not improve MAE or R² compared to using only CAP score and CAG repeat length (Tables S5 and S6 in Multimedia Appendix 1) and introduced severe multicollinearity (variance inflation factor [VIF] >12 for all features). By contrast, VIF values were acceptable when using only CAP score and CAG repeat length for both the demographic baseline model (VIF=1.2) and the final 4-feature model (maximum VIF of 1.56). This demographic baseline model was used for subsequent comparisons with xHD-Vox to quantify the specific contribution of speech features.

Automating the Selected Features’ Annotation Process Using Whisper

Whisper-extracted features and annotations-based features showed Pearson and Spearman correlations of 0.84/0.81 for numbers pronounced per second and 0.42/0.61 for the SD of the duration of pronounced numbers.

The automated pipeline achieved equal or lower MAE values than the speech pathologist-annotated pipeline (column 7 vs column 6 in Table 3), with statistically significant differences observed for TMS and TFC (paired t test; Bonferroni corrected; P<.001).

The final performance of the automated model yielded relative errors ranging from 12% for SDMT to 40% for TMS, with 13% for TFC, 19% for cUHDRS, and 35% for SW. Relative error was calculated as the MAE (column 7 in Table 3) divided by the mean cohort score (last column in Table 1).

Model Evaluation Using Longitudinal Data

Overview

First, we compared the cross-validation performance of our model obtained in (Table 3, last column) with results previously reported in the literature using comparable evaluation frameworks (Table 4), which is further discussed in the Discussion section.

Model Equations

The regression coefficients of the xHD-Vox model, calibrated on the training set, are provided in Table 5.

Model Performance on the Test Set

The R² values of xHD-Vox on the test set were 0.57 for cUHDRS and TMS, 0.49 for SW, and 0.48 for SDMT (Figure 2). The lowest R² was observed for TFC (0.26).

Compared with a demographic baseline model including CAP score and CAG repeat length, xHD-Vox improved cUHDRS prediction, increasing R² by 50% and reducing MAE by 15%. The largest gains were observed for cognitive outcomes, with relative R² increases of 81% for SW and 66% for SDMT, along with corresponding MAE reductions of 16% and 13%, respectively.

Based on the model’s cross-validation results in Table 3 (last column), the test set MAE for cUHDRS matched expectations (2.1), equal to the average MAE across the 50 folds. For Symbol Digit Modality and TMS, the test-set MAEs were lower than expected (8.2 vs 8.4 and 9.0 vs 10.1, respectively). In contrast, SW showed poorer test-set performance, with a higher MAE than during cross-validation (12.5 vs 11.6).

Mean and median absolute errors per visit remained stable over the 2-year follow-up (Figure S4 in Multimedia Appendix 1). MAE varied by HD-ISS stage (Figure S5 in Multimedia Appendix 1). For example, for cUHDRS, the MAE decreased from 2.6 at stage 0‐1 to 1.7 at stage 2 before increasing to 2.2 at stage 3.

‎

Functional Classification

Model performance for TFC prediction was modest (R²=0.26), likely reflecting the limited variance in TFC scores (mean 11.1, SD 2.4). However, speech rate was significantly lower in participants at HD-ISS stage 3 compared to earlier stages (Figure S6 in Multimedia Appendix 1; 1-way ANOVA followed by post hoc Tukey test; P<.001).

We therefore investigated whether speech rate could serve as a marker of functional impairment and trained a logistic regression classifier using CAP score and speech rate as the 2 predictors (methods and Figure S3 in Multimedia Appendix 1). The model distinguished stage 3 from stages 0‐2 with good performance on the held-out test set: recall=0.83, accuracy=0.73, and ROC AUC=0.79 (bottom right panel in Figure 2). For comparison, a model based only on CAP score achieved recall=0.68, accuracy=0.68, and ROC AUC=0.74.

Longitudinal Change

We conducted repeated-measures 2-way ANOVAs with time (Y0, Y1, and Y2) and type of measure (clinical vs xHD-Vox) as within-subject factors for all scores (Table 6). A significant main effect of time was observed for all scores (P<.001 for all scores), with no significant time × type interaction, indicating similar temporal trajectories for clinical measures and xHD-Vox predictions. Post hoc Tukey tests revealed significant differences between Y0 and Y2 for all scores, and between all time points (Y0, Y1, and Y2) for cUHDRS and SW. Mean scores decreased consistently across time points for both clinical and xHD-Vox measures, demonstrating a statistically significant longitudinal decline captured by both measures. No significant main effect of type of measure was observed. Detailed ANOVA results for the main effect of type and the time × type interaction are provided in Table S7 in the Multimedia Appendix 1. These results demonstrate that xHD-Vox reliably captures longitudinal clinical decline over a 2-year follow-up.

The magnitude of the 1-year and 2-year decline at the group level is reported in Table 7, which presents mean changes and their 95% CIs. Corresponding plots are shown in Figure S7 in Multimedia Appendix 1 for illustrative purposes. For all clinical scores, the mean 2-year decline measured by clinicians was approximately twice the 1-year decline, indicating a consistent and approximately linear progression over time. Predicted changes derived from xHD-Vox fell within the 95% CIs of the corresponding clinician-assessed changes for all scores except the 2-year change in SW. These results suggest that xHD-Vox provides a group-level estimate of the magnitude of longitudinal clinical decline.

Figure 3 compares predicted 1-year changes in cUHDRS with clinically measured changes at the individual level for both xHD-Vox and the demographic baseline model (45 observations; 3 influential outliers removed). A sensitivity analysis including all 48 observations is reported in Figure S8 in Multimedia Appendix 1 and discussed in the following paragraphs. The demographic baseline captured little intraindividual variability, with Pearson and Spearman correlations of 0.12 and 0.06, respectively, and produced identical slopes across participants.

‎

In contrast, xHD-Vox exhibited greater interindividual variability, with Pearson and Spearman correlations of 0.38 and 0.37, respectively. The model also captured positive changes in some participants at earlier disease stages (HD-ISS 0‐2), as illustrated in Figure 3 and in the individual trajectories shown in Figure S9 in Multimedia Appendix 1.

A sensitivity analysis including all 48 observations (24 participants with 2 annual change measures each) showed that correlation estimates were strongly influenced by 3 influential outliers. When all data points were included, the Pearson correlation was close to 0, whereas the Spearman correlation remained higher (ρ=0.17). The discrepancy between Pearson and Spearman correlations suggested a monotonic relationship that was disproportionately affected by a small number of extreme values, as identified in Figure S8 in Multimedia Appendix 1.

Discussion

This study presented an interpretable, automated speech-based model, xHD-Vox, for monitoring clinical progression in HD. Building on Riad et al [5] prior work using counting tasks, we reduced the feature set from 63 to 4 key predictors while maintaining comparable MAE: CAP score, CAG repeat length, and rhythm-related measures—speech rate (numbers pronounced per second) and its SD. These 2 rhythm features were automatically derived from Whisper’s transcriptions and timestamps, an open-source ASR model [8].

xHD-Vox was trained on data from 157 participants and evaluated on a longitudinal cohort of 24 participants monitored annually during 2 years. On this test set, xHD-Vox explained 57% of the variance in cUHDRS, representing a 50% improvement over the demographic baseline. The MAE was 2.2, with an ICC of 0.70—lower than the 0.92 reported for clinicians [17], but promising for a fully automated tool. Similar results were obtained for TMS (MAE=9.0; ICC=0.68 vs ICC=0.84 for clinicians [18]).

Cognitive outcomes also showed significant improvement relative to the demographic baseline, with an R² increase of 66% for SDMT, reaching 0.48. To assess clinical relevance, we compared our MAE values with normative benchmarks from Mills et al [19]. The model’s MAE for SW (12.5) approximated the difference between the 50th and 5th percentiles (11) for highly educated 50-year-olds, a demographic representative of our cohort. Similarly, the MAE for SDMT (8.2) closely matched the normative spread (8).

Although the linear regression model yielded a relatively low R² for TFC (0.26), a logistic regression using speech rate and CAP score effectively classified HD-ISS stage 0‐2 vs stage 3 with an ROC AUC of 0.79, supporting its potential utility in detecting functional decline.

Longitudinal analyses using repeated-measures ANOVAs with post hoc Tukey tests confirmed a significant decline over the 2-year follow-up in the test set for both clinician-assessed measures and xHD-Vox predictions. At the group level, the mean 1-year and 2-year changes predicted by xHD-Vox were consistent with clinically measured changes, falling within the corresponding 95% CIs. Exploratory individual-level analyses showed that xHD-Vox reasonably captured longitudinal 1-year change compared with a demographic baseline model, although results were sensitive to a small number of influential outliers. Collectively, these findings support the potential of xHD-Vox to model longitudinal clinical trajectories, while underscoring the need for larger and more diverse test samples to further assess robustness at the individual level.

Compared to prior speech studies in HD (summarized in Table 4), our model stands out for its larger cohort (n=181 gene carriers), improved performance, and longitudinal scope. In terms of performance, the MAE here was reduced for SW and TMS by 13% and 16%, respectively, relative to Riad et al [5]. For TFC, a 46% improvement was achieved compared with Nunes et al [4], although this should be interpreted cautiously, considering the difference in sample size in cross-validation (n=25 in their study). The only published longitudinal study did not model clinical outcomes but focused on the decline of speech biomarkers over time in premanifest participants (n=13) [16], which precludes direct comparison [14,16].

Unlike most previous studies, we relied on a backward counting task, a task widely conceptualized as a working memory task. Individual differences in working memory capacity are known to covary with processing speed and broader cognitive efficiency [20,21], which are core determinants of SDMT and SW performance. In addition, backward counting and the SW task both rely on inhibitory control processes: inhibition of the overlearned forward number sequence in backward counting and inhibition of the automatic reading response in the Stroop task. Thus, rhythm-related speech features likely capture these shared cognitive efficiencies, explaining their predictive value for broader cognitive performance beyond working memory alone. Speech rate was also demonstrated to be a marker of disease progression in HD [14,15]. Consistent with Vogel et al [14], we observed a reduction in speech rate in stage 3 participants compared with controls and earlier-stage patients (Figure S6 in Multimedia Appendix 1). We did not detect significant differences between stages 0‐1 and controls in our measure of speech rate, in agreement with Vogel et al [14] but in contrast to findings reported by Skodda et al [15] (refer to Table 4 for sample sizes in each group and study). Although speech rate in stage 2 participants was lower than in stages 0‐1, this difference was not statistically significant in our cohort (P=.07), whereas it was significant in Vogel et al [14]. Differences across studies may reflect variability in sample size, statistical approaches, and, importantly, the experimental task used. Indeed, Vogel et al [14] highlighted the task-dependent nature of speech rate as a biomarker, showing weak correlations between speech rate and disease burden score in a days-of-the-week recitation task (R=−0.22), but stronger correlations for reading (R=−0.68). These considerations underscore the importance of task selection when using speech rate as a cognitive and disease-related marker.

Our model can be implemented in mobile apps for remote monitoring. We provided all model equations and code on GitHub, ensured clinical interpretability through identified features, and automated feature extraction using the open-source algorithm Whisper. Whisper produced accurate transcriptions and sequence-level timestamps, resulting in high agreement with speech pathologists’ annotations for speech rate. Agreement was lower for the SD of numbers pronounced, which relies on word-level timestamps, but this limitation does not substantially compromise overall model performance. Here, the use of Whisper addressed reproducibility concerns raised by a previous study [4] that relied on proprietary ASR systems or performance issues of acoustic-based automated speech analysis [13]. Despite being trained on general speech, Whisper has demonstrated strong performance for dysarthric speech, including in alpha-synucleopathies (Pearson correlations ranging from 0.6 to 0.94 depending on linguistic features) [22]. Fine-tuning Whisper for dysarthric speech (eg, CrisperWhisper [23]) or leveraging its audio encoder opens opportunities for other predictive approaches [24].

XHD-Vox opens the door for several applications. First, it could be used in underserved areas. About 40 % of individuals with Parkinson disease currently do not receive care from a specialist [25]. Given that HD is a rare disease, we anticipate that the situation may be similar or even worse. Like Parkinson disease, HD specialists are concentrated in urban medical centers, while patients are geographically dispersed, making travel to these centers more difficult due to disability and age. By offering a simple and ecological approach for tracking disease progression, the need for clinic visits may decrease, whereas increasing the number of people benefiting from appropriate monitoring. Second, by allowing for frequent measurements of disease severity, xHD-Vox could assist clinicians in following patients, especially during crises such as COVID-19, when patients could not attend their annual visits. Third, our approach could be used for screening patients’ disease severity for clinical trials, especially between stages 0‐2 and stage 3, using our classifier.

However, xHD-Vox has not yet been tested in other languages. Fahed et al [26] conducted the only multilingual study (English, Spanish, and Polish) on acoustic biomarkers in HD, focusing on reading tasks, syllable repetition, and sustained vowels. They showed that all temporal features are language-dependent in reading tasks. In Parkinson disease, the articulation rate also differed across languages [27]. With Whisper supporting 96 languages, adapting xHD-Vox to a new language would require fewer than 100 participants to achieve 95% accuracy (Figure S2 in Multimedia Appendix 1). While we have tested our model on an unseen sample (test set) with good results, replication in cohorts from other centers would strengthen generalizability. Additionally, testing real-world data collection by recording patients on their mobile phones, as done for Parkinson disease [28], would validate the approach for remote monitoring. Building on this framework, we have developed an Android prototype of xHD-Vox, currently under internal testing, to explore its potential for mobile deployment and remote monitoring.

In conclusion, we provide novel insights into the robustness of an automated speech-based clinical estimation model for longitudinal monitoring of HD gene carriers, xHD-Vox. Its strengths lie in its automation and interpretability, facilitating deployment on mobile platforms and acceptance by clinicians. This work represents a step forward in telemedicine for HD monitoring, with the potential to detect disease progression, particularly in regions with limited medical access or during periods of health care disruption, such as lockdowns.