Use of Machine Learning to Identify Markers of Risk for Fragile X‐Associated Tremor/Ataxia Syndrome: A Preliminary Analysis

Abstract

Objective

The objective of this study was to examine whether machine learning has the capacity to prospectively identify and predict the emergence of Fragile X‐associated tremor/ataxia syndrome (FXTAS) among male fragile X premutation carriers (PCs).

Methods

We explored neuropsychological and motor evaluation metrics, brain magnetic resonance imaging (MRI), and health metrics in 103 male participants (72 PCs, mean = 60.4 years at enrollment) and 31 healthy controls (HCs; mean = 57.8 years at enrollment) across a total of 299 visits to identify optimal FXTAS risk markers. We compared different machine learning model and feature selection method combinations to identify the best features and models for (a) identifying patients with FXTAS and (b) for predicting which individuals were likely to later develop FXTAS in the study to date. Using an optimal set of features (including age, psychological symptoms, executive function and motor measures, IQ, body mass index (BMI), and structural brain measurements), we developed random forest binary classifiers for the 2 tasks. We split the dataset randomly into multiple different train and test splits and observed the average classification performance metrics across all the splits.

Results

The models showed promising ability to identify and pre‐emptively predict the emergence of FXTAS and achieved a reasonable balance between precision and recall. Accumulation of body fat (BMI), executive function weaknesses, slower reaction time and dexterity, and mental health changes, are clinical factors that may significantly increase a carrier's risk. Structural brain MRI measurements significantly added to the predictive power of the models.

Interpretation

These results suggest that machine learning has the potential to inform prediction of risk for FXTAS early, enabling better planning, timely interventions, and provision of necessary care. ANN NEUROL 2026;100:36–47

The fragile X premutation is a common trinucleotide repeat expansion (CGG) at the Fragile X Messenger Ribonucleoprotein 1 gene (FMR1) locus on the X chromosome, occurring in approximately 1 in 110 to 300 females and 1 in 250 to 850 males worldwide (or at least 850,000 people in the United States). ¹ Whereas the female “carriers” can transmit an expanded mutation to their descendants, causing Fragile X syndrome (FXS), the past 2 decades of research has revealed that carriers of the FMR1 premutation are at increased risk of neurodegenerative disease (fragile X‐associated tremor/ataxia syndrome [FXTAS]), reproductive health problems in female carriers (fragile X‐associated premature ovarian insufficiency [FXPOI]), and a wide range of additional mental and general health conditions that are still being characterized. ² , ³ , ⁴ FXTAS, the most studied and best described premutation‐associated disorder, is a progressive disease characterized by intention tremor, cerebellar gait ataxia, peripheral neuropathy, parkinsonism, and cognitive decline, and is usually accompanied by gray and white matter changes and general atrophy in the brain, and ubiquitin‐positive intranuclear inclusions upon autopsy. ⁵ It occurs in 8 to 16% of female premutation carriers (PCs) and over 60% of male PCs by age 70 years, ⁶ with highly variable penetrance in terms of onset, severity, and course, ranging from subtle and stable tremor through old age to onset in the 50s with declining mobility, debilitating tremor, dementia and personality changes, autonomic dysfunction, and finally total caregiver dependency and premature death.

Although the FMR1 CGG repeat length is modestly correlated with FXTAS age of onset and symptom severity, ⁷ it has limited prognostic value alone. By studying the preclinical phase of the disease, as well as phenoconversion (when PCs show the first clinical signs of the movement disorder) in men, we have identified neuroradiological signs (white matter hyperintensity, brainstem, and cerebellar changes), ⁸ executive function (EF; performance‐based tests and self‐report), ⁹ , ¹⁰ , ¹¹ motor (subtle balance/gait changes and slowed upper extremity movements), ¹¹ and FMR1 isoforms, ¹² , ¹³ that can help us better understand and predict the onset of the disease.

Machine learning has been successfully applied in various studies for management of various kinds of dementia, ¹⁴ , ¹⁵ Alzheimer's disease, ¹⁶ , ¹⁷ Parkinson's disease, ¹⁸ and numerous other neurodegenerative disorders. ¹⁹ , ²⁰ Due to the promise shown by statistical methods and machine learning in management of neurodegenerative disorders, in this study, we performed a statistical analysis to identify optimal markers of risk for FXTAS. We first used machine learning techniques to identify whether an individual at any timepoint has FXTAS using the identified markers. Although a machine learning model that can identify whether an individual has FXTAS is useful, it is significantly more beneficial if a model can predict whether an individual is at risk of conversion before the conversion actually occurs, because that would allow for better symptom management, planning, and potential participation in emerging treatments. We thus also designed a study to use machine learning to predict the later onset of FXTAS.

Methods

Study Design

The protocol for this study was approved by the University of California‐Davis Institutional Review Board. Our longitudinal study, following participants every 2 to 3 years, includes 72 male PCs (mean = 60.4 years at enrollment) and 31 male healthy controls (HCs; mean = 57.8 years). Because FXTAS occurs much more often in males than females, and the study examines phenoconversion, we chose to focus on males to ensure that enough participants would be likely to develop disease over the course of the project. The distribution of the number of visits per participant across the duration is reported in Supplementary Table S1. PCs were recruited primarily through announcements shared by the National Fragile X Foundation, and by referrals from colleagues at other institutions studying children with FXS and their families for research or in the clinic. Non‐carrier controls were recruited from the Institute's research registry system, social media announcements, and presentations given to local organizations. PCs traveled to UC Davis MIND Institute in Sacramento, CA, from various locations in North America, whereas the non‐carrier control sample was recruited locally throughout Northern California. Assessments included neuropsychological and motor evaluation, brain magnetic resonance imaging (MRI), CGG repeat length, and health metrics. A primary goal of the study is to enroll at‐risk carriers without FXTAS and follow them longitudinally while tracking the emergence of the disease. All participants were assigned FXTAS diagnostic determination (no, possible, probable, or definite) at each visit. We formulated the problem of predicting FXTAS as a binary classification task. We achieved this by converting the FXTAS diagnostic determination into a binary variable with “no” and “possible” values being assigned a “0” label, and “probable” and “definite” being assigned “1.” Our dataset contains 27 participants (37.5% of PCs) with at least one FXTAS positive (binary diagnosis of 1) record and 76 participants with only FXTAS negative (binary diagnosis of 0) records.

We first designed a study (study 1) to identify whether each measurement for a patient is FXTAS positive or not. In this study, the learning space for our models include measured metrics for each participant at each visit and their corresponding binary variable value indicating whether each measurement is FXTAS positive. Across all participants, the study 1 cohort has 299 total visits. There are 59 FXTAS positive data points and 240 FXTAS negative data points and the exact counts of each level of the raw FXTAS diagnosis variable are also available in Table 1.

TABLE 1.

Distribution of FXTAS Diagnosis Levels in the Dataset

Binary Label	FXTAS Diagnosis	Count	Binary Label Counts
0	0	184	240
	1	56
1	2	28	59
	3	31
Total		299

FXTAS = Fragile X‐associated tremor/ataxia syndrome.

We designed a second study (study 2) to perform early prediction of later FXTAS emergence. The goal of this study was to predict eventual conversion of FXTAS in participants based on their initial visits, when they were FXTAS negative. For this problem formulation, we first dropped all participants with “probable” or “definite” FXTAS diagnosis at their first visit. We then modified the learning space for our models by assigning the FXTAS diagnosis binary value at the latest recorded visit for each participant to every previous record for the corresponding participant in the dataset. This transformation resulted in a dataset where each record has the different metric values for each participant at each visit, but has the FXTAS diagnosis value for the latest visit of the participant, thus introducing a temporal shift between the recorded metrics and the diagnosis. We also only considered up to the first 2 visits for each patient and dropped all visits beyond the second. We performed these transformations in order to ensure that models trained on this dataset are forced to learn solely from measurements taken before participants received a binary diagnosis of 1, or when they first received a binary diagnosis of 1, thus preventing biased model performance due to information leakage from data collected from later visits when participants are already known to have converted to a diagnosis of “probable” or “definite.” Due to the removal of participants with FXTAS positive diagnosis at initial visits and retention of only up to 2 visits per participant, the cohort for study 2 has 171 data points from 92 participants. Thirty records corresponded to eventual FXTAS conversion, whereas the remaining 141 records belonged to participants who did not experience eventual conversion.

Feature and Model Selection

To obtain the most optimal markers of risk for FXTAS, we compared statistical feature selection, Mutual Information ²¹ , ²² and Recursive Feature Elimination (RFE). ²² , ²³ Identifying a smaller subset of informative features helps build simpler models that are less prone to overfitting and generalize better when introduced to new data. The process of statistical feature selection involved comparing the distributions of the values of each feature for the positive class data points to the distributions of the values of the same feature for the negative class data points using the 2‐sample Kolmogorov–Smirnov test ²⁴ , ²⁵ at a 99% significance level. For both Mutual Information and RFE, we retained the top 15 features.

Participants from study 1 were randomly split into 2 groups at an 80 to 20 ratio and the datapoints of the corresponding participants formed training and test sets, respectively. Each feature selection method was applied to the training set to derive a subset of features, each of which were then used to fit random forest, ²⁶ extreme gradient boost (XGBoost), ²⁷ logistic regression, ²⁸ and support vector machine ²⁹ models. In each case, the best model was derived based on 5‐fold cross‐validation and hyperparameter tuning. These models were then used to evaluate the corresponding test sets in order to provide an indication of which machine learning algorithms are the most promising and which feature selection algorithms produce a set of features that lead to good classification performance. The classification performance of each machine learning model and feature selection algorithm combination is reported in Table 2. Once we identified the optimal machine learning and feature selection algorithms, we applied them to the cohort for study 2 as well.

TABLE 2.

Comparing Feature Selection and Machine Learning Model Combinations

Model	Feature Selection	Accuracy	Precision	Sensitivity	Specificity	NPV	AUROC	AUPRC
RF	SFS	0.8	0.5	0.42	0.89	0.86	0.83	0.44
RF	MI	0.8	0.5	0.5	0.87	0.87	0.85	0.51
RF	RFE	0.81	0.56	0.42	0.91	0.86	0.81	0.44
LR	SFS	0.8	0.5	0.08	0.98	0.81	0.64	0.37
LR	MI	0.8	0.5	0.08	0.98	0.81	0.63	0.37
LR	RFE	0.8	0.5	0.08	0.98	0.81	0.58	0.33
XGB	SFS	0.8	0.5	0.42	0.89	0.86	0.77	0.56
XGB	MI	0.83	0.58	0.58	0.89	0.89	0.8	0.48
XGB	RFE	0.85	0.62	0.67	0.89	0.91	0.8	0.57
SVM	SFS	0.8	0	0	1	0.8	0.63	0.43
SVM	MI	0.83	1	0.17	1	0.82	0.6	0.42
SVM	RFE	0.83	1	0.17	1	0.82	0.61	0.42

Test set performance of different machine learning algorithms trained with features retained by different feature selection methods.

The figures in bold in Table 2 represent both RF and XGB, classification performance is most balanced at a threshold of 0.4. The bold values highlight the best performing combination of ML algorithm and Feature Selection Method. It shows that XGB with RFE performs the best in the preliminary arbitrary split of data, indicating tree‐based models might be the best family of models to investigate.

AUPRC = area under the precision‐recall curve; AUROC = area under the receiver operating characteristic curve; LR = logistic regression; MI = mutual information; NPV = negative predictive value; RF = Random forest; RFE = recursive feature elimination; SFS = statistical feature selection; SVM = support vector machines; XGB = Extreme Gradient Boosting.

We used an arbitrary train‐test split of the dataset to compute SHapley Additive exPlanations (SHAP) scores on the holdout split and ranked the features by SHAP scores. Supplementary Figure S1 indicates that beyond the top 15 features, as per the SHAP scores, the remainder provides diminishing marginal contributions to model performance. We thus chose 15 as the number of features to select when applying RFE to the cohorts of both studies. Figure 1 highlights the features considered by the algorithm for the cohorts of study 1 and study 2. The bars represent the frequency of retention of each feature across the 30 runs of RFE, underscoring that the top 15 features are consistently deemed informative across all train‐test splits.

FIGURE 1.

Retention frequency of features for study 1 and study 2 cohorts produced by applying RFE to the 30 splits in each cohort. Refer to Supplementary Table S2 for the complete feature names and descriptions. BMI = body mass index; BRIEF = Behavior Rating Inventory of Executive Function; RFE = Recursive Feature Elimination; RTI = CANTAB 5 Choice Reaction Time; SCL = Symptom Checklist 90 Global Severity; WMH = white matter hyperintensity. [Color figure can be viewed at www.annalsofneurology.org]

Additionally, we computed SHAP scores of all features for the cohorts of both studies and computed the mean scores for each feature across the 30 splits in each study. We compared the top 15 SHAP scores in both studies with our RFE based features and observed that, for study 1, the 2 feature sets are identical, and, for Study 2, the 2 sets differ only by a single feature (SHAP considered anxiety/mood disorder to be more important than Symptom Checklist 90 Global Severity). This further reinforces the importance and predictive power of the features selected for the final models. The mean SHAP scores are presented in Figure 2. To quantify uncertainty in feature importance, we repeated model training and SHAP attribution across 30 independent train‐test resamples for both studies. For each split, SHAP values were computed on the held‐out test data and mean absolute SHAP values were aggregated per feature. We report the mean and standard deviation of SHAP scores across resamples in Supplementary Table S5.

FIGURE 2.

Mean SHAP scores of features averaged across the 30 splits in the cohort for study 1 and study 2. Refer to Supplementary Table S2 for the complete feature names and descriptions. BMI = body mass index; BRIEF = Behavior Rating Inventory of Executive Function; RTI = CANTAB 5 Choice Reaction Time; SCL = Symptom Checklist 90 Global Severity; SHAP = SHapley Additive exPlanations; WMH = white matter hyperintensity.

Train/Test Splitting

Splitting a cohort into a single train and test set pair can by random chance result in a grouping of the data such that training and evaluation of models provides overly optimistic or overly pessimistic estimates of classification performance. Whereas we used the single random train‐test split to identify the types of models and feature selection algorithms that show promise, we designed a more rigorous experiment to assess overall model performance for our classification tasks. We divided the participants for study 1 as well as study 2 into 30 distinct train‐test splits at 80 to 20 ratios. We divided our cohorts at the participant level in order to ensure that measurements from the same participant do not appear in both training as well as test sets, resulting in data leakage into the models.

Averaging the classification performance metrics of models evaluated on 30 different test sets and computing the confidence intervals for each metric provides a more robust assessment of the capabilities of our classifiers than the performance of a model on a single random train‐test split.

Model Development and Evaluation

In study 1, we developed models to classify each measurement of each participant as FXTAS positive or not FXTAS positive. We applied the optimal feature selection method to each of the 30 training sets for study 1 and selected the 15 most frequently retained features across the 30 sets of selected features. We then used the best performing machine learning algorithm to fit a model with these 15 features on each of the same 30 training sets. We performed 5‐fold cross‐validation and hyperparameter tuning and selected the optimal model in each case. These 30 models were then evaluated on their corresponding holdout sets and the accuracy, precision, sensitivity, specificity, negative predictive value (NPV), area under the receiver operating characteristic curve (AUROC), and area under the precision‐recall curve (AUPRC) for each were measured. The mean and 95% confidence interval of the 30 values of each of these metrics were computed. The process of training, evaluation, and averaging was performed thrice, each time with a different classification threshold (0.4, 0.5, and 0.6). The results of these experiments are reported in Table 3.

TABLE 3.

Classification Performance of Models in Study 1

Model	Threshold	Accuracy	Precision	Sensitivity	Specificity	NPV	AUROC	AUPRC
Random forest	0.4	0.85 ± 0.02	0.67 ± 0.07	0.61 ± 0.07	0.91 ± 0.02	0.91 ± 0.02	0.87 ± 0.03	0.69 ± 0.07
	0.5	0.85 ± 0.02	0.72 ± 0.08	0.50 ± 0.06	0.94 ± 0.02	0.89 ± 0.01
	0.6	0.85 ± 0.02	0.78 ± 0.08	0.38 ± 0.07	0.97 ± 0.02	0.87 ± 0.01
XGBoost	0.4	0.84 ± 0.02	0.65 ± 0.08	0.56 ± 0.08	0.91 ± 0.02	0.9 ± 0.02	0.83 ± 0.07	0.63 ± 0.07
	0.5	0.84 ± 0.02	0.68 ± 0.08	0.47 ± 0.08	0.93 ± 0.02	0.88 ± 0.02
	0.6	0.84 ± 0.02	0.65 ± 0.12	0.37 ± 0.09	0.95 ± 0.02	0.87 ± 0.02

Average classification performance across 30 test sets in study 1 for Random forest and XGBoost models evaluated at different classification thresholds.

AUPRC = area under the precision‐recall curve; AUROC = area under the receiver operating characteristic curve; NPV = negative predictive value; XGBoost = Extreme Gradient Boosting.

The cohort for study 2 has a class imbalance due to the removal of FXTAS positive first visits from the dataset, resulting in a further reduction of the minority class. In order to mitigate this, after performing feature selection and obtaining the 15 most informative features (similar to study 1), we upsampled the minority class in each of the 30 study 2 training sets using Adaptive Synthetic Sampling (ADASYN) ³⁰ and balanced the ratio of positive and negative classes. We used the best performing machine learning algorithm to fit models on the upsampled training sets with the 15 most informative features. Like we did in study 1, we performed 5‐fold cross‐validation and hyperparameter tuning, and selected the best model for each training set. We evaluated these 30 models on the corresponding test sets and computed the accuracy, precision, sensitivity, specificity, NPV, AUROC, and AUPRC. The mean and 95% confidence intervals of each of these metrics were computed across the 30 splits, and the process of training, testing, and averaging was performed for classification thresholds of 0.4, 0.5, and 0.6. The results of these experiments are reported in Table 4.

TABLE 4.

Classification Performance of Models in Study 2

Model	Threshold	Accuracy	Precision	Sensitivity	Specificity	NPV	AUROC	AUPRC
Random forest	0.4	0.88 ± 0.02	0.78 ± 0.10	0.52 ± 0.08	0.96 ± 0.02	0.9 ± 0.02	0.88 ± 0.03	0.71 ± 0.07
	0.5	0.88 ± 0.02	0.71 ± 0.15	0.41 ± 0.10	0.98 ± 0.01	0.89 ± 0.02
	0.6	0.85 ± 0.02	0.64 ± 0.17	0.27 ± 0.09	0.98 ± 0.01	0.86 ± 0.02
XGBoost	0.4	0.86 ± 0.02	0.66 ± 0.09	0.44 ± 0.08	0.95 ± 0.02	0.89 ± 0.01	0.8 ± 0.05	0.56 ± 0.08
	0.5	0.86 ± 0.02	0.67 ± 0.11	0.41 ± 0.08	0.96 ± 0.02	0.88 ± 0.02
	0.6	0.86 ± 0.02	0.65 ± 0.14	0.34 ± 0.10	0.97 ± 0.01	0.87 ± 0.02

Average classification performance across 30 test sets in Study 2 for Random Forest and XGBoost models evaluated at different classification thresholds.

The figures in bold in Table 4 represent both RF and XGB, classification performance is most balanced at a threshold of 0.4.

AUPRC = area under the precision‐recall curve; AUROC = area under the receiver operating characteristic curve; NPV = negative predictive value; XGBoost = Extreme Gradient Boosting.

The most informative features for study 1 and study 2 are available in Table 5.

TABLE 5.

Selected Features

Study 1 Features	Study 2 Features
Age	Age
Symptom Checklist 90 Global Severity	Behavioral Dyscontrol Scale
CANTAB Spatial Working Memory	Symptom Checklist 90 Global Severity
CANTAB Stop Signal Test	CANTAB Spatial Working Memory
CANTAB Rapid Visual Information Processing	CANTAB Stop Signal Test
CANTAB Paired Associates Learning	CANTAB Rapid Visual Information Processing
CANTAB Choice Reaction Time	CANTAB One Touch Stockings of Cambridge
Full Scale IQ	CANTAB Choice Reaction Time
Behavior Rating Inventory of Executive Function (MCI)	Behavior Rating Inventory of Executive Function (MCI)
Behavior Rating Inventory of Executive Function (BRI)	Behavior Rating Inventory of Executive Function (BRI)
Purdue Pegboard Test	Purdue Pegboard Test
Body Mass Index	Body Mass Index
Brainstem Corrected for TCV	Brainstem Corrected for TCV
Cerebellum Corrected for TCV	Cerebellum Corrected for TCV
WMH Corrected for TCV	WMH Corrected for TCV

Top 15 most informative features for study 1 (left) and study 2 (right) according to RFE. For each study, these are the 15 most frequently retained features across the 30 sets of 15 features produced by applying RFE to the 30 training splits.

MCI = mild cognitive impairment; RFE = Recursive Feature Elimination; WMH = white matter hyperintensity.

We also tested the hypothesis that brain structure features are the most informative set of features by applying the modeling methodology for both study 1 (with and without upsampling minority classes) and study 2 but with brain structure features removed. The 15 features most frequently retained after applying the optimal feature selection method to the 30 training sets of the cohorts from both studies were utilized to fit models on the training sets using the best performing machine learning algorithm. For study 1, models were fit once without upsampling minority classes, and then after upsampling minority classes (Supplementary Table S4). For study 2, minority classes were upsampled before fitting any models. The models were then evaluated on the corresponding holdout sets at classification thresholds of 0.4, 0.5, and 0.6. The means and 95% confidence intervals of the 3 sets of classification performance metrics are reported in Supplementary Table S4.

In order to assess the calibration of our models, for both the studies, we calculated Brier scores ³¹ for model performances on each of the holdout sets, and, as baseline, we computed Brier scores for uninformative models on the same holdout sets that assigned all samples the same probability equal to the outcome prevalence in the training data. For both studies, the model Brier scores and baseline Brier scores were averaged across the 30 splits, and percentage improvement over baseline was computed.

Results

Study 1

Applying different combinations of feature selection methods and machine learning algorithms to the random training split and evaluating on the corresponding test split revealed XGBoost model fit with features retained by RFE produces the best performance (see Table 2). However, because random forest models appeared to perform comparably, we implemented both these tree‐based algorithms for our subsequent experiments.

We compared the average performance of random forest and XGBoost models (averaged across 30 holdout sets) fit with the 15 most informative features produced by RFE. We also compared the average performance of the models at classification thresholds of 0.4, 0.5, and 0.6. Although the random forest and XGBoost models seem to perform comparably, the average performance of the random forest models across the 30 holdout sets slightly exceeds that of the XGBoost models for each of the 3 classification thresholds. Additionally, we also observed that both random forest and XGBoost models provide the most balanced performance at a classification threshold of 0.4. Higher thresholds produce better precision but at the expense of sensitivity. The best performance overall is observed using random forest models at a classification threshold of 0.4. The classification performance metrics for these experiments are listed in Table 3.

The mean baseline Brier score was 0.16 and the mean model Brier score was 0.11, indicating a 31% improvement in probabilistic accuracy relative to the baseline. The standard deviation in model Brier score was 0.032.

Study 2

The 15 most informative features produced by RFE across the 30 training splits for study 2 are listed in Table 5. Random forest and XGBoost models fit on the upsampled training sets using those features were evaluated on the holdout sets at classification thresholds of 0.4, 0.5, and 0.6, and the average performance metrics revealed a trend similar to that of study 1. Once again random forest models outperformed XGBoost models at most metrics across all thresholds. For both models, 0.4 was the optimal value of classification threshold in this study as well, with the random forest with a threshold of 0.4 performing the best. However, unlike study 1, increasing the classification threshold did not result in a noticeable improvement in precision at any point and only worsened sensitivity. The classification performance metrics for these experiments are presented in Table 4.

The mean baseline Brier score was 0.15 and the mean model Brier score was 0.11, resulting in a 27% improvement in probabilistic accuracy relative to the baseline. The standard deviation in model Brier score was 0.035.

Importance of Brain Structure Features

The average classification performance (across 30 train‐test splits) of random forest models fit using the 15 most informative features identified by RFE after brain structure features were removed from the cohorts are reported in Supplementary Table S4. We observed that in the absence of brain features, the models struggled to identify FXTAS positive measurements and were especially poor at identifying early conversion of FXTAS. Thus, brain structure features are integral to our models and hold the most predictive power among all groups of features considered in our experiments.

Discussion

One of the most pressing questions for fragile X PCs is whether and when they may develop FXTAS. This question has critical implications for each individual as they prepare for and experience the aging process, but also for clinical researchers as we attempt to identify those at highest risk and optimal candidates for prophylactic treatments and lifestyle adjustments that may slow, or perhaps even prevent the emergence of the disease. Currently, whereas there are no markers that can be used to confidently predict FXTAS conversion, ongoing studies have identified several potential preclinical changes that appear to precede the movement disorder. In this study, we leveraged existing longitudinal data from a cohort of male PCs to determine whether machine learning has the potential to identify the most promising early markers of FXTAS. Although machine learning has been used effectively to search for medical conditions that may be over‐represented in FMR1 carriers, ³² , ³³ this is the first time that it has been applied to the study of neurodegeneration in this group.

Whereas our study currently has several important limitations, discussed below, the results provide evidence that certain clinical, physical health, and structural brain variables may provide important prognostic information for the later emergence of FXTAS in male PCs. Beyond an individual's premutation status (only carriers can develop FXTAS) and age, these preliminary results suggest that accumulation of body fat (body mass index [BMI]), executive function weaknesses, slower reaction time and dexterity, and mental health changes, are clinical factors that may significantly increase a carrier's risk. However, it is especially notable that structural brain MRI measurements significantly added to the predictive power of the models. This is notable because a brain MRI is usually not indicated for a PC until the clinical signs of FXTAS emerge, at which time these neuroradiological markers (white matter hyperintensities and gross volume loss) are used to aid in diagnosis. The present results may emphasize the need for earlier radiological monitoring in adult PCs well before any clinical signs arise, so that lifestyle changes and other targeted treatments can be provided as they become available. Before this can be applied in practice, it will be necessary to establish methods that can be feasibly used by clinical radiologists in a cost‐effective manner, with empirically derived, clinically significant cutoff metrics applicable to individual patients. Once established, the movement disorder neurologist might then use probabilistic risk data derived from the empirical studies to communicate to the patient and their family, and perhaps to guide treatment. These metrics may also be useful for enriching clinical trial cohorts, as prophylactic interventions might then be tried in PCs at highest risk for imminent phenoconversion.

Although our models show promising capability of identifying risk of FXTAS in individuals, our studies currently have several limitations. Although the proportion of PCs in our sample that met diagnostic criteria for probable or definite FXTAS was somewhat lower than what has been reported in screening studies, ⁶ perhaps limiting its representation, we specifically recruited men without FXTAS to enter the study in order to study the phenoconversion process. Because of this critical study design criterion, the proportion we report should not be interpreted or reported as an estimate of prevalence. In addition, it is unclear whether the present findings would be replicated in a study of female patients as the neurodegenerative process in women with the premutation may differ significantly from men. ³⁴ For example, in one study, the middle cerebellar peduncle white matter hyperintensity that is included as a major radiological sign in the diagnostic criteria for FXTAS occurred in only 11% of female PCs compared to up to 60% of male PCs in previous reports. ⁵ We used upsampling to mitigate the severe class imbalance that exists in our dataset, but the models would potentially benefit from being able to learn from more real positive class data points, especially in study 2 for early prediction of conversion. Because the models perform relatively well despite being limited by the learning space constraints, it is reasonable to anticipate that the models would demonstrate superior classification performance if provided with more data from a larger study. Indeed, a substantially larger sample of PCs studied over a longer time horizon will provide more comprehensive and accurate information about factors contributing to FXTAS conversion. A larger sample size and additional timepoints will also allow for inclusion of additional potentially important predictors. For example, recent studies suggest that educational attainment may buffer individuals with the premutation, especially those with mid‐range CGG repeats, from developing motor or cognitive changes (Klusek et al, 2026) ³⁵ and even mortality, ³⁶ , ³⁷ and others generally from neurodegenerative diseases (eg, see Ref. 38). Whereas classification on a 4‐level diagnostic scale collapsed into a binary outcome provides a clinically interpretable first step, ordinal or survival modeling approaches on a larger sample with additional predictors may better capture disease trajectories and heterogeneity in progression timing. Multi‐center collaborations and studies of other cohorts are also needed to provide external validation and scalability of the approach.

We formulate our studies as binary classification problems based on an FXTAS diagnosis variable with 4 discrete levels that we convert into a binary target variable. Although we treat positive labels as “FXTAS positive,” it is difficult to identify a discrete threshold for differentiating between “FXTAS positive” and “FXTAS negative.” Thus, utilizing an FXTAS diagnosis target variable with more discrete levels or a continuous variable for the same would allow for designing more sophisticated studies for prediction of conversion.

Despite these limitations, we identified optimal markers for risk of FXTAS and using these markers, we were able to identify individuals who have FXTAS and also pre‐emptively predict conversion of FXTAS. We obtained the best predictive performance when using tree‐based models in conjunction with RFE. Between random forest and XGBoost models, the former were found to be superior, and we tuned the classification threshold to obtain the best balance between precision and recall. Using these methods, we were able to demonstrate the potential of machine learning techniques for identifying potentially important features for identifying FXTAS and also for prediction of onset of FXTAS, facilitating timely planning, care, and necessary interventions.

Author Contributions

C.G., A.P., S.M.R., R.H., and D.H. contributed to the conception and design of the study; C.G., A.P., J.Y.W., C.C., A.S., F.T., G.E., and J.F. contributed to the acquisition and analysis of data; C.G. and D.H. contributed to drafting the text or preparing the figures.

Potential Conflicts of Interest

Nothing to report.

Supporting information

Table S1. Participant Counts for Different Numbers of Visits.

Table S2. Descriptions of all the Features in the Dataset.

Table S3. Classification Performance with Minority Class Upsampling in Study 1.

Table S4. Classification Performance Without Brain Structure Features.

Table S5. Mean and Standard Deviation (STDEV) of SHAP Feature Importance Across 30 Resampled Train‐Test Splits.

Figure S1. SHAP Scores of Features Ranked in Descending Order for an Arbitrary Train‐Test Split of the Dataset.

Data Availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.