Huntington's disease (HD) is a rare, genetic, neurodegenerative disease caused by a cytosine adenine guanine (CAG) repeat expansion variant of the huntingtin gene (HTT) (1) and is characterised by a triad of cognitive, behavioural, and motor symptoms (2, 3). Disease onset, defined as the onset of motor signs and symptoms as measured by a Diagnostic Confidence Level of 4 (3, 4), typically occurs in the prime of life, between the ages of 30 and 50 years (2). HD is associated with increasing disability, worsening of function and loss of independence, leading to death within approximately 15 years of onset (2, 5). Motor and cognitive symptoms deteriorate steadily as the disease progresses (3, 6–9), while behavioural symptoms tend to be episodic (10).

With potential disease-modifying therapies for HD in clinical development (11), there is interest in measuring disease progression and characterising prognostic variables in order to improve the efficiency and accuracy of clinical trials (12). Prognostic variables can be used to identify a patient population through an enrichment strategy to reduce interpatient variability in clinical trials or alternatively to enrich for faster progressors, and to eventually inform the optimum time to start treatment (12). Statistically controlling for prognostic baseline variables may also be important in non-randomised (e.g., open-label) studies as they could confound the relationship between treatment exposure and outcomes. Additionally, when testing hypotheses in randomised studies, the probability of detecting a treatment effect will usually increase by including prognostic variables as covariates in the analysis, as this would explain a significant amount of variability observed due to random error.

Large prospective cohort studies have shown that manifestations of progression, that is, clinical signs and symptoms of HD, as well as known biological predictors of progression such as CAG repeat length and CAG-age product (CAP) score, can predict clinical progression or motor onset (7, 13). However, no study has systematically ranked the importance of predictors of progression in a manifest HD population (i.e., after the onset of unequivocal motor symptoms).

Random forest (RF) regression models permit interrogation of large, complex clinical datasets to capture non-linear associations between multidimensional predictive variables and clinical outcomes with high predictive accuracy (14, 15). RF approaches are well-suited to classification and regression problems, such as identifying variables with predictive potential for disease progression from clinical datasets. Here, we use modern machine learning methods to examine a large number of HD variables to identify the most important predictors of progression on five clinical outcomes: total functional capacity (TFC), a measure of function; stroop word reading (SWR), a measure of attention and psychomotor processing speed; symbol digit modalities test (SDMT), a measure of executive function, visuo-spatial working memory, attention and processing speed; total motor score (TMS), a measure of motor function; and the composite unified HD rating scale (cUHDRS), an equally weighted composite outcome measure of the TFC, TMS, SDMT, and SWR that was developed based on an early manifest HD population (16). The large prospective Enroll-HD cohort (NCT01574053) is used to investigate the relative contribution and ranking of potential prognostic variables to predict clinical progression in a clinical trial-like manifest HD population.

The analysis included 1,608 individuals meeting typical criteria for clinical trials in manifest HD and with CAG repeats between 36 and 64 (filtering criteria shown in Table 1). Patient demographics are shown in Table 2.

The highest-ranked variables predictive of disease progression for each outcome are shown in Figure 1 and the top 10 variables for each outcome shown in Table 3. CAP was found to be the most predictive variable for all outcomes and CAG repeat length was ranked as the second most important variable for all outcomes. Other prognostic variables associated with faster progression that ranked in the top 10 for at least three of the five outcomes were: age at diagnosis (all but SWR and TFC), being accompanied to clinic visits (for all outcomes), history of cognitive impairment (all but SWR), tetrabenazine use (for all outcomes) and antipsychotics use (all but TMS). The effect of these variables on disease progression trajectory as measured by the cUHDRS is shown in Figure 2.

The common variables among the top 10 most important features for all outcomes were: CAP score, CAG repeats, accompanied or unaccompanied at clinic visit, tetrabenazine use, antipsychotics use and having severe cognitive impairment.

Unadjusted R² measures were calculated for the RF models including CAG and CAP score only and compared with the model including all the features (Table 4). Using additional features with CAP and CAG can capture the variance of outcome by 17% more for cUHDRS and 15% more on average for the other outcomes. The slight improvement in model fit with CAP, CAG and age compared with the model built with the shared top 10 features could be due to the different cross-validation sets, and also the very high contributions of CAP and CAG to the model fit.

This analysis used real-world data from the large Enroll-HD registry and a machine learning algorithm to identify novel predictors of HD progression with significant impact on the slope of clinical decline observed over a 2-year follow-up period. The two most important predictors identified were CAP score and CAG repeat length, in agreement with previous studies (7, 13). In addition, several strong predictors were identified that have either not been previously studied (being accompanied to a visit) or have had inconsistent effects in other studies (cognitive impairment, use of tetrabenazine or antipsychotics) (7, 17–19). The novel variables identified were predictive of progression over multiple clinical domains, measured by motor, cognitive and functional endpoints, as well as the composite endpoint. Using all predictors in addition to known prognostic variables improved the ability of the model to predict clinical outcomes (see video abstract in the Supplementary Materials).

Some of the features tested, which may have been expected to rank highly as prognostic variables based on previous studies in premanifest HD (i.e., prior to the onset of unequivocal motor symptoms)—including smoking, alcohol intake and body mass index (BMI) (20–22)—were not found to be important predictors of progression. These results may not be directly comparable to the current study, which was carried out in a manifest HD population. It is also known that self-reporting of smoking and alcohol use is unreliable in the general population, as revealed by studies using advances in DNA methylation measurement to assess substance use status (23). We found BMI to be a weak discriminating factor among patients with different values of change in outcome. A further potential explanation for the disparity between our findings and previous studies could be that the prognostic value of these variables may be dependent on disease stage. In the current study, the population was relatively progressed, and it is possible that other variables associated with the disease could outweigh environmental variables.

It should be noted that we used an RF algorithm with the setting that prevents bias in the ranking based on the data structure. Whilst it is still possible that a feature can rank highly due to collinearity with another feature that is a strong predictor of the outcome, this could be prevented by calculating the conditional importance of the features, which is computationally very complex (24).

Additionally, the observed associations are based on observational data and are therefore not indicative of causal relationships, due to measured and unmeasured potential confounding factors. For example, being accompanied to clinic visits may affect the clinical outcome scores measured by virtue of the companion's additional report which informs the clinical rating. It may also be because healthier participants are able to continually attend visits alone, whereas those who are on worse clinical trajectories need additional emotional or practical assistance (e.g., driving) to complete visits. Similarly, antipsychotics may be used to treat motor symptoms in HD, and therefore may be expected to reduce TMS without influencing overall disease progression. Cognitive outcome measures may also be related to variables including dementia and severe cognitive impairment. RF approaches have good performance in modelling complex, multidimensional disease-specific datasets (like Enroll-HD) (25). In this application, an RF approach was used to find novel associations (e.g., identify variables with predictive potential for disease progression), and does not imply causality (i.e., the aetiological role of the variable during disease progression) (26). Nevertheless, by virtue of the strength of the associations observed, some of these features may be important to control for in analyses of observational studies and may have implications for companion participation in interventional trials.

The current study focused on clinical variables only and did not include imaging or fluid biomarkers, which previous studies have suggested may be predictive of disease progression (7, 27, 28). This limitation was due to the nature of the currently available HD databases. In this study, we used the Enroll-HD database, which provides a sufficiently large sample size for the analysis but does not include imaging or biofluid data as part of the main study. Imaging databases such as PREDICT-HD (NCT00051324) are available, but do not provide the comprehensive range of clinical variables that is available in Enroll-HD, such as medication history. The available biofluid databases, such as the HD-CSF study (28, 29), are too small to be informative on the scale of the current analysis.

A further limitation is that the results described here are based on a selected cohort intended to reflect the inclusion criteria of ongoing clinical trials, and therefore may not be representative of the wider HD population, including younger patients (juvenile-onset HD), elderly patients (>65 years), late-stage patients (>Stage III) or premanifest patients. Further research is needed to determine the wider applicability of these results to these populations.

This study made use of a supervised RF regression model to identify putative and novel predictors of disease progression in HD. Identifying prognostic variables usually requires large sample sizes to optimise predictive accuracy, which may be a limitation for rare conditions such as HD. Since 2012, the Enroll-HD registry, which includes over 19,000 participants from 177 sites in 20 countries, has allowed a large, high-quality dataset to be available for researchers to advance the understanding of HD. The power of RF modelling is particularly relevant within the context of HD, where improved understanding of this multidomain disease and need for efficient trial design is evident.

To overcome known methodological limitations, RF approaches are being developed to harness the full potential of long-term registry data in clinical risk prediction and may, in future, accelerate disease risk and course prediction in HD. Given the dynamic nature of disease, the recently published RF Survival, Longitudinal and Multivariate model was developed to evaluate the temporal nature of variables (such as rate of change of variables) (30). Such approaches will further refine identification of clinically meaningful predictive variables not only for risk of disease progression as a static entity, but risk of disease progression over time and clinical course. Such temporal approaches will prove useful for future studies in HD, where disease course is highly variable.

In summary, the RF approach described here using the Enroll-HD dataset has identified novel prognostic variables which may be important candidates for statistical control in clinical trials and observational studies in HD.

The data analysed in this study was obtained from Enroll-HD, https://enroll-hd.org/, the following licenses/restrictions apply: to access data you must be a researcher employed by a recognized academic institution, company or non-profit organisation and apply for an Enroll-HD access account. Requests to access these datasets should be directed to https://enroll-hd.org/for-researchers/become-a-qualified-researcher/.

All the codes and algorithms for the analysis are available upon request.

NG contributed to the conception and design of the study, and acquisition and interpretation of data for the work. SS contributed to conception of the study, and acquisition and interpretation of data for the work. GP and PW contributed to the design of the study. JL contributed to the conception and design of the study. RH contributed to conception and design of the study, and acquisition and analysis of data for the work. All authors drafted or substantively revised a significant portion of the manuscript or figures, and approved the final version for submission.

NG, RH, SS, and GP are employees of F. Hoffmann-La Roche Ltd. JL is a paid Advisory Board member for F. Hoffmann-La Roche Ltd and uniQure biopharma B.V, and a paid consultant for Vaccinex Inc, Wave Life Sciences USA Inc, Genentech Inc and Triplet Inc. The remaining author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The authors declare that this study received funding from F. Hoffmann-La Roche Ltd. The funder was involved in the study design, analysis, interpretation of data and the decision to submit it for publication.