Objective Tinnitus Measurement via fNIRS and Machine Learning

Tinnitus affects around six to 20 percent of adults, with approximately 20 percent of these having severe and debilitating symptoms, such as depression and cognitive dysfunction.1, 2 Despite its wide prevalence, no clinically objective test is available to determine the presence or severity of tinnitus or assess the effectiveness of treatments.3 Developing an objective measure of tinnitus will enable a more accurate assessment of this condition, marking an important step in the development and assessment of potential treatments.Shutterstock/sdecoretFigure 1: fNIRS Montage. Sixteen sources (red circles) and 16 detectors (gray circles) forming channels were placed on the frontal, temporal, and occipital regions of the scalp. Channel numbers are shown. 36 long and 4 short channels (marked by * and yellow source-detector links) were formed. Audiology, machine learning, artificial intelligenceTo address this concern, we used functional near-infrared spectroscopy (fNIRS), which monitors changes in blood oxygen levels, allowing the imaging of the oxygen demands of active brain regions. fNIRS is non-invasive and non-radioactive. It operates quietly, features a good temporal resolution, and has several advantages over other imaging modalities used to localize human brain areas that show changes in activity related to tinnitus.2, 4 It is also portable and cost-effective, which are important for routine clinical use. Changes in the level of spontaneous neural activity, neural synchrony, and reorganization of cortical tonotopic maps have been correlated with tinnitus.2 Several cortical regions identified in tinnitus neuroimaging studies are accessible using fNIRS, including the auditory cortex, frontal cortex, and cuneus.4-6 Most reported findings in brain imaging studies on tinnitus are based on a statistical analysis of either resting state or evoked brain activity.2 Machine learning methods are well suited to integrating signal features from multiple channels and different conditions with providing a diagnosis for a single subject rather than showing group-level differences as statistical methods usually do. Subjective ratings of tinnitus severity can be used as input to machine learning algorithms to map fNIRS signal features to severity levels (training phase). fNIRS features from other individuals can then be classified to tinnitus severity levels based on past observations. In this study, we applied statistical and machine learning algorithms to fNIRS signals to (1) assess the sensitivity of fNIRS to differentiate individuals with tinnitus from controls and (2) identify fNIRS features associated with subjective ratings of tinnitus severity and whether these could differentiate between perceived loudness of tinnitus and annoyance. We first performed statistical analysis to gain a better understanding of signal features, cortical regions, and conditions that show group differences and changes with tinnitus severity levels, and to avoid our machine learning algorithms becoming a ‘black box’. STUDY METHODS fNIRS recordings were performed on 25 participants with chronic subjective tinnitus (23 experience it bilaterally) and 21 healthy adults with no history of tinnitus, neurological, or hearing disorders. There were no significant group differences in age or hearing thresholds. Data from three healthy participants were excluded due to signal quality or technical issues with the fNIRS. Tinnitus severity was assessed using the Tinnitus Handicap Inventory (THI),7 which quantifies the perceived severity of tinnitus and associates score ranges with different severity levels (e.g., 0 to 16 for slight tinnitus, 58 to76 for severe). Participants with tinnitus also rated the loudness and annoyance of their tinnitus on a scale of 1 to 10 before each recording. fNIRS signals were collected using a continuous-wave fNIRS system (NIRScout, NIRx Medical Technologies LLC), with 16 sources and 16 detectors placed over the frontal, temporal, and occipital cortical regions (Fig. 1). Each source-detector pair (called a ‘channel’), was placed ~30mm apart. Using source-detector pairs placed close together (~11mm), systemic signals from superficial layers were recorded and used to remove systemic artifacts from other channels. The fNIRS test session consisted of three recording periods. The first recording period was a six-minute resting-state recording with participants sitting still with their eyes closed. The two remaining periods recorded fNIRS signals in response to 15-second auditory or visual stimuli presented at random, and 20- or 25-second non-stimulus intervals in between. Several preprocessing steps were applied to the fNIRS signals to remove channels with poor signal quality from further analysis. Signals from the remaining channels were converted to optical density and concentration changes of oxygenated and deoxygenated hemoglobin (HbO and HbR) estimated using the modified Beer-Lambert law.8 fNIRS RESTING-STATE STATISTICAL ANALYSIS & FINDINGS Seed analysis was used to investigate resting-state functional connectivity networks by finding correlations between seed regions and other brain regions (e.g., see 9) using whitened correlations.10 In this study, a left seed was used by averaging channels 9 and 10 (Fig. 1), and a right seed was obtained by averaging channels 30 and 31 (channels estimated to cover the superior temporal and Heschl's gyrus). HbO and HbR correlation values for channels comprising frontal and occipital regions of interest (ROI) were then averaged for statistical analysis. Frontal channels were located over the superior frontal gyrus, medial, superior frontal gyrus, medial orbital, and middle frontal gyrus. The occipital channels covered the cuneus and superior occipital gyrus. Connectivity measures between both left and right seeds with frontal HbO signals were higher in the tinnitus group, with right seed differences reaching significance. This was not found for left seed connectivity or HbR signals. In the tinnitus group, HbO-derived connectivity between left and right seeds and frontal channels increased with the duration of tinnitus with the correlation on the right side approaching significance. Right seed- occipital connectivity values derived from HbR signals were significantly higher in the tinnitus group. This was not found for left seed connectivity. In the tinnitus group, HbR-derived connectivity between the right seed and occipital channels increased significantly with subjective ratings of loudness. fNIRS EVOKED RESPONSE STATISTICAL ANALYSIS & FINDINGS To analyze evoked responses, HbO and HbR signals were epoched from t = –5 to t = 30s relative to stimulus onset and epochs with excessive amplitudes due to noise rejected. Mean HbO and HbR amplitudes across time windows 0 to 5 seconds (for auditory responses) and 10–15 seconds (visual) were calculated. These time windows were chosen based on the initial rising phase of the group averaged responses and waveform morphology.11, 12 fNIRS evoked response amplitudes were averaged over ROIs for statistical analysis. Left and right temporal ROIs included channels on the left and right sides covering the superior, middle and inferior temporal gyrus, Heschl's gyrus, and angular gyrus. Visual evoked responses were averaged over the occipital ROI, covering the cuneus and superior occipital gyrus. There was no significant difference between left and right auditory responses. Averaged across both sides, the HbO auditory response was smaller in the tinnitus group. This result could be due to the increased background neural activity present in tinnitus leading to saturation of the hemodynamic response.13 Visual response amplitudes were significantly larger in the control group. COMBINING fNIRS FEATURES USING MACHINE LEARNING To combine features from HbO and HbR resting state and evoked response signals from fNIRS channels over different cortical regions, machine learning methods including feature selection and classifiers were used. Features input to these algorithms included auditory and visual response amplitudes and frontal and occipital connectivity measures described above. Features from all channels were used, and Information Gain, a feature selection algorithm, was used to rank features based on their importance in classification. These features were then used with four different classification methods to classify participants as controls or experiencing tinnitus. Classifiers were also used to differentiate patients with tinnitus as having slight/ mild versus moderate/ severe tinnitus (based on THI ratings). The four classifiers used were Naïve Bayes, K-nearest neighbor (KNN), Rule Induction, and Artificial neural networks (ANN). Classifier performance was assessed using only connectivity measures, only evoked response features, or using both connectivity and evoked features to assess the relative importance of the different features. To calculate the performance of these algorithms, 10-fold cross-validation was used and the average sensitivity (true positive rate), specificity (true negative rate), and accuracy (number of correctly predicted samples over the total number of samples) were calculated. The best accuracy in differentiating tinnitus participants from controls was achieved using auditory features alone and a Naïve Bayes classifier, resulting in an accuracy of 78.3 percent. The highest accuracy of 87 percent for differentiating slight/ mild (n = 18) from moderate/ severe (n = 7) tinnitus was achieved using connectivity measures with the Neural Network classifier although the sensitivity obtained was low (51.23%). FUTURE DIRECTIONS Our statistical findings support previous research that has identified measures of brain connectivity or evoked responses associated with tinnitus. We have built on these findings further using fNIRS and machine learning, and have shown the feasibility of this approach to classify an individual's fNIRS data to a tinnitus or control group as well as a tinnitus severity level with promising accuracy. Confirmation of findings with larger sample sizes is needed to validate our models. Tinnitus, by nature, will always have a subjective component. However, an objective measure will help measure certain aspects of tinnitus that will assist with the development and testing of new treatments.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.