Objective Assessment of Progression and Disease Characterization of Friedreich Ataxia via an Instrumented Drinking Cup: Preliminary Results

A Bayesian Network Approach for Friedreich Ataxia Severity Classification using Probability Modelling

Friedreich ataxia (FRDA) requires an objective measure of severity to overcome the shortcoming of clinical scales when applied to trials for treatments. This is hindered due to the rarity of the disease resulting in small datasets. Further, the published quantitative measures for ataxia do not incorporate or underutilise expert knowledge. Bayesian Networks (BNs) provide a structure to adopt both subjective and objective measures to give a severity value while addressing these issues. The BN presented in this paper uses a hybrid learning approach, which utilises both subjective clinical assessments as well as instrumented measurements of disordered upper body movement of individuals with FRDA. The final model's estimates gave a 0.93 Pearson correlation with low error, 9.42 root mean square error and 7.17 mean absolute error. Predicting the clinical scales gave 94% accuracy for Upright Stability and Lower Limb Coordination and 67% accuracy for Functional Staging, Upper Limb Coordination and Activities of Daily Living.Clinical relevance- Due to the nature of rare diseases conventional machine learning is difficult. Most clinical trials only generate small datasets. This approach allows the combination of expert knowledge with instrumented measures to develop a clinical decision support system for the prediction of severity.

Prediction of the disease course in Friedreich ataxia

We explored whether disease severity of Friedreich ataxia can be predicted using data from clinical examinations. From the database of the European Friedreich Ataxia Consortium for Translational Studies (EFACTS) data from up to five examinations of 602 patients with genetically confirmed FRDA was included. Clinical instruments and important symptoms of FRDA were identified as targets for prediction, while variables such as genetics, age of disease onset and first symptom of the disease were used as predictors. We used modelling techniques including generalised linear models, support-vector-machines and decision trees. The scale for rating and assessment of ataxia (SARA) and the activities of daily living (ADL) could be predicted with predictive errors quantified by root-mean-squared-errors (RMSE) of 6.49 and 5.83, respectively. Also, we were able to achieve reasonable performance for loss of ambulation (ROC-AUC score of 0.83). However, predictions for the SCA functional assessment (SCAFI) and presence of cardiological symptoms were difficult. In conclusion, we demonstrate that some clinical features of FRDA can be predicted with reasonable error; being a first step towards future clinical applications of predictive modelling. In contrast, targets where predictions were difficult raise the question whether there are yet unknown variables driving the clinical phenotype of FRDA.