A novel machine learning approach for the classification of Huntington disease manifestation using motor features.

BACKGROUND

Currently, diagnosis of manifest Huntington disease (HD) is made when there is ≥ 99% confidence that observed motor abnormalities, rated on the Unified Huntington Disease Rating Scale (UHDRS), are due to HD (Diagnostic Confidence Level, DCL, = 4). However, there are no explicit rules relating UHDRS motor score items to an appropriate DCL, leaving diagnosis potentially subjective and variable.

OBJECTIVE

We sought to develop algorithms to classify manifest HD, based on UHDRS motor scores.

METHODS

We applied supervised machine learning techniques including elastic net logistic regression, random forest and our novel "robust support vector machine" (rSVM) designed to handle data uncertainty, to build three different classification models for predicting HD class (manifest/non-manifest) based on motor scores. Existing DCL-based classifications formed the reference standard. Models were trained and tested using data from two, large observational studies - Enroll-HD and COHORT.

RESULTS

All three models demonstrated ≥95% classification accuracy and ≥90% sensitivity and specificity for the diagnosis of manifest HD across both study populations. 47.8% of Enroll-HD cases classified as manifest by the rSVM but non-manifest by the DCL-based reference were considered manifest by another clinical classifier, Participant Category. The majority of these people also had high DCLs and many had high total motor scores so may reflect inconsistencies or data entry errors in the DCL classifications.

CONCLUSIONS

We have developed three highly accurate models for diagnosing manifest HD using UHDRS motor scores. This has potential applications as clinical decision support tools and for data checking and standardization in clinical trials.