Related Experiment Video
Updated: Sep 27, 2025
A Machine Learning Approach to Design an Efficient Selective Screening of Mild Cognitive Impairment
Published on: January 11, 2020
The Rapid Screening for Somatosensory Tinnitus Tool: a Data-Driven Decision Tree Based on Specific Diagnostic
Sarah Michiels1,2,3, Emilie Cardon2,4,3, Annick Gilles2,5,4
1REVAL Rehabilitation Research Center, Faculty of Rehabilitation Sciences, Hasselt University, Diepenbeek, Belgium.
Background:
Somatosensory or somatic tinnitus (ST) is a type of tinnitus where changes in somatosensory afference from the cervical spine or temporomandibular area alter the tinnitus perception. Very recently, the diagnostic value of a set of 16 diagnostic criteria for ST was determined. The next step in the development of easily applicable diagnostic criteria is to provide an uncomplicated model, based on the existing criteria, which can easily be used in clinical practice.
Objectives:
This study aims to construct an accurate decision tree, combining several diagnostic criteria, to optimize both sensitivity and specificity of ST diagnosis.
Design:
An online survey was launched on the online forum Tinnitus Talk, managed by Tinnitus Hub in a convenience sample of participants with tinnitus. The survey included 42 questions, both on the presence of diagnostic criteria for ST and on other potentially influencing factors. A decision tree was constructed to classify participants with and without ST using the rpart package in R. Tree depth was optimized during a five-fold cross-validation. Finally, model performance was evaluated on a subset containing 20% of the original dataset.
Results:
Data of 7981 participants were used to construct a decision tree for ST diagnosis. Four criteria were included in the final decision tree: 'Tinnitus and neck/jaw pain increase/decrease simultaneously', 'Tension in suboccipital muscles', 'Somatic modulation', and 'Bruxism'. The presented model has an accuracy of 82.2%, a sensitivity of 82.5%, and a specificity of 79%. Receiver operator characteristic curves demonstrated an area under the curve of 0.88.
Conclusions:
Based on a 42-item survey, a decision tree was created that was able to detect ST patients with high accuracy (82.2%) using only 4 questions. The RaSST is therefore expected to be easily implementable in clinical practice.
