Random Forest Classification to Predict Response to High-Definition Transcranial Direct Current Stimulation Therapy for Tinnitus

Authors: Emilie Cardon12, Laure Jacquemin12, Griet Mertens12, Paul Van de Heyning12, Olivier M. Vanderveken12, Vedat Topsakal12, Vincent Van Rompaey12, Annick Gilles123
1Department of Translational Neurosciences, Faculty of Medicine and Health Sciences, Campus Drie Eiken, Antwerp University, Antwerp, Belgium
2Department of Otorhinolaryngology – Head and Neck Surgery, Antwerp University Hospital, Edegem, Belgium
3Department of Education, Health and Social Work, University College Ghent, Ghent, Belgium

Background Noninvasive brain stimulation techniques such as transcranial direct current stimulation (tDCS) are hypothesized to reduce tinnitus severity by modifying cortical hyperactivity in brain regions associated with tinnitus perception. However, individual response to tDCS has proven to be extremely variable. We employed random forest (RF)
classification, a supervised machine learning approach, to predict individual treatment response in a large cohort of chronic subjective tinnitus patients.

Methods Ninety-nine tinnitus patients received six biweekly sessions of high-definition (HD) tDCS of the right dorsolateral prefrontal cortex (rDLPFC). Before treatment and at a three-month follow-up time point, patients filled in the Tinnitus Functional Index (TFI), a Visual Analogue Scale (VAS) to assess tinnitus loudness and the Hospital Anxiety and Depression Scale (HADS). Handedness was assessed using the Edinburgh Handedness Inventory (EHI). RF classification with five-fold cross-validation was applied to predict whether an individual responded positively to the treatment, defined as a reduction of at least 13 points on the TFI.

Results Thirty out of 99 patients responded positively to HD-tDCS treatment. The RF classifier predicted treatment response with an accuracy of 82.41% and an area under the curve of 0.902, corresponding to a sensitivity of 80.95% and a specificity of 84.71%. Feature importance analysis revealed that patients with higher TFI, VAS and HADS depression scores and negative EHI scores, corresponding to left-handedness, were more likely to respond positively to HD-tDCS treatment.

Conclusion Individual response to HD-tDCS treatment may be accurately predicted using easily obtainable questionnaire data. Patients with higher tinnitus burden, concurrent depressive symptoms and/or a preference of the left hand may respond more positively to HD-tDCS of the rDLPFC. Although larger studies are necessary to validate the proposed classifier, our results might provide the basis for better and more personalized guidance towards the right treatment for chronic tinnitus patients.