Seventy-one young adults participated in this study and were split into three groups: a preliminary group of healthy participants without tinnitus to optimize the silent-in-noise detection task, a tinnitus group, and a matched control group (Fig. 2A). Male and female participants between 18–30 years of age, in good health other than tinnitus were recruited from posts on the University of Oxford and Oxford Experiment Group on Facebook and invited to participate in the study. Inclusion criteria were age (18–30 years), normal or corrected-to-normal vision, no diagnosed health conditions other than tinnitus, and availability for 12 months after starting the study. Exclusion criteria were any major medical or psychiatric condition (other than tinnitus), hearing loss (defined as $>$25 dB hearing level (HL) in both ears), severe hyperacusis (assessed using the Khalfa hyperacusis questionnaire [23, 24]), claustrophobia, or belonging to a vulnerable group.

A general questionnaire was used to assess age, gender, current educational status, current educational area, previous education, free school dinner eligibility, languages spoken in order of familiarity, highest music grade and the instrument on which it was obtained, whether participants were claustrophobic or dyslexic, history of hearing problems, medical conditions and any current medications. Tinnitus sufferers answered additional questions which assessed their history of tinnitus, duration of chronic tinnitus, whether the tinnitus is pulsatile and previous or ongoing tinnitus therapy.

The Edinburgh Handedness Inventory (EHI; [25]) was used to assess the participant’s predominant hand and the Khalfa Hyperacusis Questionnaire (HQ; [23, 24]) was used to assess hyperacusis. Tinnitus severity was assessed by using the Tinnitus Functional Index (TFI; [9]) and Tinnitus Handicap Inventory (THI; [10]). In addition, if individuals with tinnitus perceived their tinnitus as a hearing disability, we assessed the significance of any hearing disability using the Hearing Handicap Inventory (HHI; [26]).

The first group of eighteen individuals without tinnitus (preliminary group) was allocated to examine the effect of the position and saliency of the silent gap within the stimulus sound on the performance of the operant silent gap in-noise detection paradigm (preliminary experiment). The other participants (n = 53) were a tinnitus group (n = 20) and a matched control group (n = 33) used in a case control observational study. Tinnitus participants were accepted if they had chronic tinnitus ($>$3 months) that was not pulsatile (following a heartbeat rhythm), not objective (noise that can be heard by an external examiner or created internally in the body) or a consequence of a medical disorder. Tinnitus participants were not accepted if they were undergoing any tinnitus therapy at the same time as our study.

In the preliminary group, only operant silent gap-in-noise detection was tested following questionnaires and audiogram. The optimized silent gap-in-noise detection paradigm was then tested in the tinnitus and control groups, alongside silent GPIAS and ABRs (Fig. 2A).

The tests were organized into blocks of 100 trials each, where the duration and position of gaps were randomized whilst the type of sounds (30 kHz low pass filtered broadband noise (BBN), and one octave narrowband noises (NBN) either centred at 1, 2, 4, 8 or 16 kHz) and gap amplitude modulations (100%, 80%, 60%) were kept fixed in each block. To maintain subject focus and attention, testing was spread over several one hour-long sessions with regular breaks. It took between 3 and 5 hours to complete the testing. This was predominantly carried out in a single day for each participant, although on occasions was performed over two days when necessary due to tiredness. All testing was conducted in a double-walled sound-attenuating chamber (IAC acoustics, Winchester, UK; dimensions 1020 wide, 1250 deep and 2000 mm high).

Audiometry testing followed the standard audiometry guidelines (2018) from the British Society of Audiology [27], updated to include testing at 16 kHz. Pure tone bursts (0.25, 0.5, 1, 2, 4, 6, 8, 16 kHz) at different intensity levels (generated using an RP2.1 real-time processor, 100 kHz sampling rate) and attenuated using a PA5 programable attenuator (TDT, Alachua, FL, USA) were presented to each ear separately through headphones (HD 250, Sennheiser, Wedemark, Germany) to extract hearing thresholds. Any frequency threshold above 25 dB transformed HL resulted in exclusion from the study. Two participants from the control group failed to meet this criterion. Tinnitus participants self-matched their tinnitus percept at each ear to a computer-generated pure tone presented over headphones by varying the frequency (0.25 to 16 kHz) and intensity of the pure tone (0–90 dB sound pressure level, SPL). One participant reported hearing a “white noise” tinnitus and was excluded from the study. Another participant with pulsatile tinnitus was also excluded.

Participants were exposed to sound stimuli (400 ms duration, 75 dB SPL) from a speaker (TW026MO, Audax, AAC, La Chartre Sur le Loir, France) positioned level with their heads. The stimuli used were 30 kHz low pass filtered BBN, and one octave NBN either centred at 1, 2, 4, 8 or 16 kHz (computed in MATLAB v2022, MathWorks, Natick, MA, USA) and generated using a real-time processor RP2.1 (TDT, Alachua, FL, USA), sampling rate 100 kHz, and amplified by an SA8 power amplifier (TDT, Alachua, FL, USA). In half of the trials, a silent gap was inserted pseudo-randomly, either at 100, 200 or 300 ms from the start. Using a two-alternative forced-choice paradigm, participants pressed the left button to indicate that they had heard a silent gap and the right button to indicate that they heard a continuous sound (Fig. 2B). The task events, stimulus presentation and responses were controlled by custom scripts written in MATLAB controlling the TDT hardware (RP2.1 and SA8).

Testing was performed in blocks of 100 trials where the silent gap duration was randomized whilst the type of stimulus was fixed (BBN or 5 different NBN one octave narrowband noises either centred at 1, 2, 4, 8 or 16 kHz). Six independent blocks of 100 trials each were performed by each participant. In the preliminary group, the depth of the gap was modulated at 100%, 80% and 60% intensity attenuation (named mod 0 at 100%, mod 2 at 80% and mod 4 at 60% intensity modulation). This was repeated in independent testing blocks for each stimulus type and gap modulation combination (18 blocks in total: 6 types of stimulus and 3 gap depth modulations). For each of the 18 combinations, the duration of the silent gap was varied at either 1, 3, 4, or 10 ms between trials.

In the control case observational study, both control and tinnitus groups were presented with a wider range of silent gaps (1, 3, 5, 10, 20, 50, 100 and 270 ms), the position of the silent gaps in each stimulus was randomly varied in each trial and 80% modulation depth was used for all gap trials. Blocks of 100 trials were used for each stimulus type combination (6 blocks in total).

For each participant, eleven supervised training trials were conducted before data acquisition to allow familiarization with the equipment and sounds presented. To prevent a testing learning effect over time, the order of frequencies and modulations tested was randomized [28, 29] in each participant from the three experimental groups. To guide them during testing, a monitor displayed the trial number and a reminder of what each button meant: the left display box read “two noises” and the right display box read “one noise” (Fig. 2B). These labels were used instead of “gap” and “no-gap” to reduce the cognitive load.

The eye blink reflex (Preyer reflex) in response to a startle stimulus was measured using the surface electromyographic signal (sEMG) of the left orbicularis oculi muscle [30]. Surface electrodes (AG/AgCl non-metallic radiolucent 10 mm diameter disposable EEG disc electrodes, filled with Ten20 conductive Neurodiagnostic Electrode Paste, Weaver & Co, Aurora, CO, USA) were placed on the right side of the face and attached to the skin using skin-safe tape. To minimize movement artefacts and standardize attention across participants, participants were invited to play ‘Where’s Wally?’ during testing. Electrical signals were amplified and digitised (25 kHz) by a low impedance headstage (RA4LI-RA16PA) connected to a RA16 and controlled by Synapse software (v.95, TDT, Alachua, FL, USA).

Participants were presented with 40 startle stimuli, 100 ms BBN 120 dB SPL, on a background of continuous 60 dB SPL BBN [31]. These parameters typically elicit a blink response [32], limit habituation [33], and minimize boredom or discomfort. In half of the trials, a 50 ms silent gap was introduced within the background noise, 150 ms before the startle stimulus. These gaps were incorporated pseudo-randomly. The interval between each trial was also pseudo-randomly allocated between 6–17 s to reduce predictability [31, 32, 33, 34].

We used the logarithmic acoustic startle reflex ratio as a proxy measure of pre-pulse inhibition (PPI, following Ornitz [35]).

$$\text{PPI}=1-\left(\frac{\text{Startle Amplitude in a gap trial}}{\text{Startle Amplitude in a no}-\text{gap trial}}\right)$$

Using rodent models of tinnitus, Schilling et al. (2017) [34] demonstrated a statistical approach to measuring PPI. They created a vector of all possible combinations of gap and no-gap trial amplitude ratios (ASR). Through trialling several transformations, they found logarithmic ASR, log(ASR), to be Gaussian-like distributed. We bootstrapped log(ASR) for each individual with 2000 re-samples to obtain a distribution of log(ASR) means. We used the mean of this distribution as the proxy measure for participant pre-pulse inhibition.

To exclude unrelated blinking, we only considered EMG responses which occurred 21–120 ms after stimulus presentation [31] and excluded responses with an amplitude lower than the mean of the baseline plus 2 standard deviations. Response latency was taken as the time between stimulus onset and maximum EMG amplitude within the response window.

Using headphones, participants were presented with clicks of alternating polarity at a rate of 89.9 Hz and pure tones (2 and 4 kHz). Stimuli were presented 2000 times in 4 blocks initially of 0, 15, 30, and 45 dB HL. However, due to difficulties in isolating individual peaks, the HLs were later increased to 75, 80, 85, and 90 dB HL. The stimulus presentation rate was 27.7 Hz. Sentiero equipment and software (v2.5.1, Path Medical, Germering, Germany) were used to produce the stimuli. ABRs were recorded using surface electrodes (DORMO SX-30, Ag/AgCl E.C.G./E.K.G., Telic Group, Bigues i Riells, Spain) with Signa gel (Parker Laboratories, Fairfield, NJ, USA), a highly conductive multi-purpose electrolyte electrode gel if the impedance was lower than 6 k$\mathrm{\Omega }$. Electrodes were placed on the right side of the head.

Data analysis was performed with MATLAB (MathWorks), and R (version 4.3, R Foundation for Statistical Computing, Vienna, Austria). All data were analysed using custom-written scripts in MATLAB. We tested for normality (Shapiro Wilk [SW] test when n 50 and Kolmogorov-Smirnov [nKS] test when n $>$50) and according to the outcome we either used parametric (Levene’s Test for homoscedasticity followed by analysis of variance (ANOVA) and post-hoc Bonferroni test) or non-parametric tests (Kruskal-Wallis [KW] test followed by either Mann-Whitney U [MWU] or Kolmogorov-Smirnov [KS] tests) for group comparisons. $\chi$² tests of proportions were used to assess minimal clinically important difference in TFI scores, and Pearson correlation coefficients were used to relate changes in TFI scores to changes in tinnitus duration. Two-sided p 0.05 indicated statistical significance. ABRs peak amplitude and latency values extracted by Sentiero software were exported for analysis in MATLAB.

For operant silent gap-in-noise detection in the preliminary experiment the sensitivity index, d’, was calculated to adjust the hit rate (HIT) for false alarms (FA):

$$d^{\prime }=z(HIT)-z(FA)$$

where z(x) is the inverse normal distribution z-score of the proportion x (HIT or FA). Values of d’ were adjusted when HIT was equal to 1 or FA was equal to 0; in these instances, d’ was estimated as being 1–(1/2T) or 1/2L, respectively, where T is the number of targets and L is the number of lures. Subject accuracy was plotted against gap length to produce psychometric function curves. Gap detection threshold was defined as the gap length when d’ = 1, corresponding to approximately 60% correct responses in the task. The slope of the psychometric function at threshold (Fig. 3A) was also used as a measure of sensitivity. Optimal performance was defined by the maximum value of d’ and the lapse rate (the proportion of incorrect responses at the longest gap length (10 ms), for which high values may indicate a lack of attention). The statistical significance of the results was tested using multilevel ANOVA with subject identity (ID) as a random factor. Results are reported as mean $\pm$ standard error (SE). To assess the statistical significance in gap detection thresholds and slopes, we used a bootstrap procedure (200 repeats) in which d’ values were calculated after resampling the trials to produce a range of psychometric functions for each stimulus type, which formed a reference distribution. From this distribution, we then performed statistical analysis using multilevel ANOVA.

In the observational study that included control and tinnitus participants, due to adding longer gap durations ($>$10 ms, maximum 270 ms), fitting of the psychometric function using d’ was less reliable than using the left button press probability. Therefore, the latter metric was used to generate the psychometric curves (Fig. 3B), to which we fitted sigmoid functions. This is equivalent in signal detection theory to plotting the hit rate for silent gap trials and the false alarm rate for no gap trials (example shown in Fig. 3). Performance was assessed by the gap detection thresholds (GDT) defined as (1) the gap duration at which the participant’s hit-rate was 65% (65% GDT) and (2) the gap duration at the psychometric inflection point (the point on a curve where the gradient changes sign - ipGDT), both as estimators of their gap detection threshold; (3) the maximum asymptote as an indication of maximum performance; (4) the slope gradient to estimate sensitivity; (5) and the false alarm rate for no gap trials to quantify attention/bias.

Performance of the silent gap-in-noise detection paradigm for durations greater than 100 ms was poorer than predicted. This could be explained by both inattention and pre-emptive button pressing. Therefore, data for gap-in-noise of durations greater than 100 ms were only included if the performance exceeded those for silent gap-in-noise durations of 50 ms and less. Participant data was also excluded if their FA rate was over 50%, indicating inattention and/or bias towards detecting silent gaps.

In the preliminary group experiment (n = 18), both gap position (100, 200, 300 ms) and attenuation of the sound intensity during the gap (depth modulation 100% (mod 0), 80% (mod 2) and 60% (mod 4)) were varied randomly between individual trials with the aim of decreasing saliency and predictability (Fig. 1B). Salience was found to depend on depth modulation, with salience increasing with depth modulation. Silent gap detection thresholds increased when salience and predictability of the gap were reduced (Fig. 3A and Fig. 4). No differences were found in the reaction time (ranging between 800 and 900 ms) regardless of the modulation depth used, the position of the gap, or the type of stimulus (BBN or one octave NBN centred at 1, 2, 4, 8, 16 kHz).

Due to limited data, we used hit and lapse rates, rather than d’ prime, to estimate performance according to gap position (Fig. 3A, Table 1). We found that for gap lengths $\ge$3 ms, hit rate was significantly higher when the gap was in the middle compared to the beginning of the stimulus (Table 1). Post-hoc comparison revealed that gaps at the end of the stimulus had the lowest hit rates (mixed-effects ANOVA, p 0.01). However, the position of the gap had no effect on hit rates (proportion of correct responses) when the gap was as short as 1–2 ms, suggesting that this is close to the limit of detection.

Lapse rates followed the same pattern as hit rates according to performance. For all stimulus types, gaps in the middle of the stimulus produced the lowest lapse rates, gaps at the beginning produced significantly higher lapse rates and gaps at the end produced the highest lapse rates (mixed-effects ANOVA, p 0.01; Table 1).

The reduction of gap salience through depth modulation had a significant effect on its detection indicated by the rising detection thresholds (Fig. 4A), decreased sensitivity (Fig. 4B) and increased false alarm rates (Fig. 4C). When assessing each stimulus type, gap detection thresholds were significantly higher at increased modulation values for all narrowband sounds except 8 kHz. Gap detection threshold at 4 kHz NBN, for example, increased from 5.23 $\pm$ 0.05 ms with no modulation to 6.86 $\pm$ 0.17 ms at 80% intensity modulation and further to 8.92 $\pm$ 0.04 at 60% intensity modulation (F_(2,5) = 183.55, p 0.001). The change in thresholds observed at higher modulations was greatest for 2 kHz NBN stimuli, for which threshold was 3.46 $\pm$ 0.10 ms with no modulation, 5.56 $\pm$ 0.22 ms at mod 2 modulation, and 9.23 $\pm$ 0.31 ms at mod 4 (F_(2,5) = 725.15, p 0.001). Without modulation, the gap detection threshold at 2 kHz NBN is more similar to the detection thresholds of the higher frequencies (8 kHz and 16 kHz) and BBN than the lower frequencies (Fig. 4D. F_(2,5) = 4685.1, p 0.001). The fact that 2 kHz NBN has a particularly low gap detection threshold at low modulation, but the highest at high modulation, suggests that as a frequency it is the most sensitive to the effects of modulation of the silent gap.

The slope of the psychometric function at d’ = 1 (threshold) indicates the dynamic sensitivity of the test, with higher slopes associated with a greater increase in d’ to small gap length differences (Fig. 3A). Slope depends on the spectral quality of the stimulus. As with gap detection threshold, lower frequencies (1 kHz; 2 kHz and 4 kHz NBN) were associated with flatter slopes, and higher frequencies (8 kHz, 16 kHz NBN and BBN) with steeper slopes (at d’ = 1. F_(5,199) = 495.29, p 0.001. Fig. 4E). This effect suggests that silent gaps are easier to detect in stimuli containing higher frequencies. This manifests as lower gap detection thresholds (Fig. 4D), steeper slopes (Fig. 4E) and reduced variability of false alarm rate (Fig. 4F) at higher frequencies.

Gap detection sensitivity (slope) also depends on the modulation of gap stimuli (Fig. 4B). For stimuli with higher modulation, slopes were less steep (F_(2,199) = 205.03, p 0.001) (slope at d’ = 1 [fraction], mean $\pm$ SE: mod 0: 0.44 $\pm$ 0.02; mod 2: 0.40 $\pm$ 0.01; mod 4: 0.34 $\pm$ 0.01). Where gap detection threshold values were very similar for different modulations (e.g., at 8 kHz NBN and BBN stimuli, where the range of gap detection thresholds across all modulations was as low as 0.28 ms), slope provides some discrimination in subject performance.

False alarm rates varied across individual subjects according to their internal decision criterion value. However, if subject identity is nested as a random factor, inter-individual variability can be accounted for, and differences in FA across conditions may be interpreted as a useful performance indicator. Modulation and stimulus type had a significant effect on FA (Fig. 4C,F). FA is dependent on gap modulation, with greater FA rates at higher values of modulation (FA rate [fraction], mean $\pm$ SE: mod 0: 0.19 $\pm$ 0.01; mod 2: 0.23 $\pm$ 0.02; mod 4: 0.25 $\pm$ 0.02; F_(2,17)= 6.01, p 0.01). An increase in FA across all individuals at higher modulation values suggests that for mod 2 and mod 4 stimuli (Fig. 4B), it is progressively harder to discriminate gap and no-gap stimuli.

Higher frequency stimuli were associated with lower FA rates than lower frequency stimuli (F_(5,17) = 3.53, p 0.001. Fig. 4F). This indicates that for lower frequency stimuli, subjects were more likely to guess the presence of a gap, in turn suggesting they found these harder to discriminate (Fig. 4F).

Lapse rates (the proportion of gap-containing stimuli with maximal gap length (10 ms) that the subject failed to detect) can be used to indicate subject attention, as 10 ms gaps are the easiest gaps to identify and are generally consistently above threshold values. For all stimulus types, lapse rates were higher with greater gap modulation (lapse rate [fraction], mean $\pm$ SE: mod 0: 0.10 $\pm$ 0.01; mod 2: 0.16 $\pm$ 0.02; mod 4: 0.30 $\pm$ 0.03; F_(2,17) = 4.43, p 0.01). Lapse rates at mod 4 were particularly high. Additionally, mod 4 stimuli were significantly affected by stimulus type (F_(5,7) = 5.46, p 0.01) (lapse rate [fraction], mean $\pm$ SE: 1 kHz NBN: 0.42 $\pm$ 0.07; 2 kHz NBN: 0.39 $\pm$ 0.08; 4 kHz NBN: 0.46 $\pm$ 0.07; 8 kHz NBN: 0.07 $\pm$ 0.03; 16 kHz NBN: 0.30 $\pm$ 0.04; BBN: 0.19 $\pm$ 0.08). Post-hoc comparison revealed that the lapse rates for 1 kHz NBN, 2 kHz NBN, 4 kHz NBN and 16 kHz NBN were significantly different to 8 kHz NBN and BBN at mod 4, but that stimulus types within these latter two groups were not significantly different to one another.

Although the participants were not given any instructions regarding the timing of their responses, it is thought that the response time (RT) is indicative of the responder’s confidence, and that this will vary according to the difficulty of the task. There was a significant difference in the mean RT for some subjects (F_(17,405) = 50.14, p 0.001), demonstrating the individual variability of the cohort. However, when subject ID was nested as a random factor, it revealed no significant effect of gap modulation on RT (F_(2,17) = 0.38, p = 0.68).

For most gap lengths, there was no difference in mean RT, but for 10 ms gap stimuli the mean RT was significantly lower than for other gap lengths (Supplementary Fig. 1A; F_(5,17) = 3.94, p 0.01). This suggests that subjects recognized 10 ms gaps more rapidly than stimuli with other gap lengths (including no-gap stimuli). Stimulus type had a significant effect on RT, even when subject ID was included as a nested random factor (F_(5,17) = 4.23, p = 0.001). Whilst post-hoc comparisons were not significant between the RTs for 1 kHz, 2 kHz and 16 kHz NBN, mean RT for 8 kHz NBN stimuli was lower than the RTs for 4 kHz NBN and BBN stimuli (Supplementary Fig. 1B; F_(5,17) = 4.23, p = 0.001).

Next, we discuss the results of the main study which compared a tinnitus cohort with matched-controls. Hearing status was established by audiometry and ABR measurements (see methods). Two control participants were excluded due to unilateral hearing loss and two tinnitus participants were excluded for not meeting the inclusion criteria. Between the control and tinnitus groups, there were significant differences in musical ability (Table 2) with controls having a median of 8 in their best musical grade whereas the tinnitus group had only 4 (Mann-Witney U Test, U = 755, p = 0.0014). Significant differences were also observed in the hyperacusis questionnaire with higher scores in the tinnitus group (Table 2). The modified Khalfa Hyperacusis questionnaire comprises of 14 items that evaluate someone’s sensitivity to everyday sounds within the attentional, social and emotional domains. The maximum score is 42 points. Tinnitus participants scored 39 versus 25 in controls (Table 2; Mann-Witney U Test, U = 684.5, p = 0.0018).

Following audiometry, the tinnitus percept for the tinnitus group was individually assessed in each ear. This was achieved by tinnitus participants self-matching their tinnitus percept to a computer-generated pure tone presented over headphones by varying the frequency (0.25 to 16 kHz) and intensity (0–90 dB SPL) of the pure tone (see methods). Despite matching the tinnitus of both ears independently, there was a strong intra-individual correlation between the tinnitus percept in each ear for most of the tinnitus participants (frequency r = 0.7528; intensity r = 0.6479). The median tinnitus frequency identified for tinnitus participants was 9.2 kHz for the left ear and 8.9 kHz for the right ear, with median intensity levels of 22 and 26 dB SPL respectively. Therefore, the tinnitus percept was close to the hearing threshold (19 dB for 8 kHz pure tone (ISO 389-5 part 2)) in the majority of the tinnitus cohort (Table 3).

Gap detection thresholds were estimated according two criteria: when hit rate was 65% (65% GDT) and at the inflection point in the psychometric function (ipGDT) (Fig. 3B; see methods). Both estimations, 65% GDT and ipGDT, show similar trends across the stimuli used (BBN and NBN 1, 2, 4, 8, 16 kHz; Fig. 5A) and with good correlation between them (Table 4).

All participants exhibited the same relationship between the gap detection threshold and stimulus type of the sound presented. Thresholds were overall lowest for BBN stimuli. Among NBN stimuli, thresholds increased as stimulus frequency decreased, independently of the threshold estimator used, although ipGDT had a lower variance than 65% GDT (Fig. 5, Table 4).

Overall, the controls had lower gap detection thresholds than tinnitus participants (Fig. 5; 2-way ANOVA, F_(1,24)= 12.3, p 0.01). This trend in difference was observed for all stimulus types tested except 1 kHz NBN (the lowest frequency tested). Post hoc analysis revealed that the differences were statistically significant at 2 and 8 kHz NBN (Student’s t-test, p 0.05). The differences at 2 kHz NBN for 65% GDT (6.96 vs. 6.20 ms; p = 0.032, KS) (Fig. 5C) and 8 kHz NBN for ipGDT (3.87 vs. 3.11 ms; p = 0.026, MWU) (Fig. 5D) were statistically significant.

Although the tinnitus was always chronic ($>$3 months), the duration of the tinnitus ranged from 5 to 120 months with a median of 44 months (Table 3). In the tinnitus group, we found that gap detection thresholds for NBN stimuli centred at and below 4 kHz were positively correlated with the duration of the tinnitus (Supplementary Fig. 2), suggesting that gap detection performance worsens with tinnitus duration.

Tinnitus participants completed two different questionnaires to assess the impact of their tinnitus: The TFI, with 8 domains (intrusiveness, sense of control, cognitive, sleep, auditory, relaxation, quality of life, and emotional), and the THI (Table 5). Scores from both questionnaires were highly correlated (Pearson Correlation r = 0.937) (Fig. 6A), so we calculated a combined tinnitus index (CTI) as the mean of the two questionaries’ scores. CTI scores negatively correlated with the tinnitus duration (r = –0.555. Fig. 6B), indicating that the longer the tinnitus duration the lower the impact it had on subject’s daily life.

Using CTI scores, we categorised participants into 5 groups (Fig. 6C) following the same criteria used for TFI [9]. A CTI of 0 was given to controls for comparison purposes. We found that tinnitus participants tended to have low CTI scores (slight to moderate) (Fig. 6C; Table 5). Interestingly, all TFI subcategory scores increased with CTI scores, except for the audiological subcategory (Table 5).

When comparing control and tinnitus cohorts stratified by CTI grade, those with slight tinnitus (CTI 1) had greater 65% GDTs relative to controls for 4 kHz (7.89 vs. 4.58 ms, p = 0.007, 65% GDT; 7.22 vs. 3.8 ms, p = 0.002, ipGDT; MWU) and 8 kHz NBN (5.73 vs. 3.464 ms, p = 0.002, 65% GDT; 4.8 vs. 3.11 ms, p = 0.001, ipGDT; MWU). For the same frequencies, these individuals with slight CTI 1 tinnitus also had greater 65% GDTs compared to those with and relative to those mild CTI 2 tinnitus (for 4 kHz p = 0.007, 65% GDT; MWU and 8 kHz p = 0.004, 65% GDT; p = 0.007 ipGDT; MWU) (Fig. 7A,B).

As previously mentioned, a negative Spearman’s correlation was found between CTI score and tinnitus duration (–0.555, p = 0.007). Linear regression analysis revealed that 26% of the variation in CTI could be accounted for by tinnitus duration. Longer tinnitus duration might explain the gap detection threshold impairment found in CTI 1, the group where tinnitus had the lowest functional impact measured by the questionnaires.

We assessed control and tinnitus participants’ eye blink reflex using surface electrodes to detect the contraction of their orbicularis oculi muscle in response to a startle stimulus (Fig. 8A,B). Participants for which blink responses were not recorded in $>$50% of trials were categorised as blink-fails and were excluded. Blink-fails occurred because of early habituation or a low signal-to-noise ratio that could be observed at any time during testing. The blink-fail rates for control and tinnitus groups were 26% and 27% respectively. Participants categorized as non-blink-fails and included in the analysis rarely exhibited habituation across the time course of the testing session. Although some participants from either the control or tinnitus groups presented a very stereotyped blink responses across trials, others had a much greater variability across trials. Therefore, instead of simply calculating the mean over all trials, we bootstrapped log(ASR) with 2000 samples (ASR = a vector of all possible combinations of gap and no-gap trial amplitude ratios) to generate a distribution of means for each participant. We then compared the means of each distribution. Over 50% of control participants showed pre-pulse inhibition (Fig. 8). We found no difference in mean log(ASR) between control and tinnitus groups (p $>$ 0.05, t-test. Fig. 8). There was no significant difference in the distributions of mean log(ASR) between tinnitus subgroups classified according to CTI (p = 0.06, Kruskal-Wallis). There was no significant difference in the prepulse inhibition between controls and tinnitus. However, this result is difficult to interpret since very few participants showed a consistent blink response. For example, 50% of control participants did no show pre-pulse inhibition regardless of the gap length.

Blink response typically consisted of several peaks (Fig. 8) that lasted around 150–200 ms. Latency of the first peak was around 50 ms. We analyzed the response latencies from gap and no-gap trials between control and tinnitus groups. We calculated an average response latency to gap and to no gap trials. All groups were normally distributed, so we conducted a one-way ANOVA, and found no significant difference between control and tinnitus groups for either gap or no gap trials (gap trials F_(1,41) = 1.867, p = 0.179; no gap trials F_(1,41) = 0.331, p = 0.959).

ABRs were recoded with Sentiero equipment using clicks and pure tone bursts (2 and 4 kHz tone burst (TB)) presented through headphones. Initially, we used 0, 15, 30, and 45 dB HL (Fig. 9). However, the peak ABR wave amplitudes were too low, especially wave I, so in subsequent participants we used 75, 80, 85, and 90 dB HL (Figs. 9,10).

ABR traces were similar across all three stimulus types (Clicks, 2, and 4 kHz TB). Typical traces can be found in Fig. 9. Three waves (I, III and V) were identifiable in most participants, with waves becoming more obvious from I to III to V. All stimulus types elicited a predictable increase in wave latency as stimulus intensity decreased. Invalidity rates for each stimulus type, when the trace was not good enough to extract data, were similar overall, at 10.7%, 8.3% and 7.4% for 2 kHz, 4 kHz TB and Click stimuli respectively.

The amplitude (Pearson’s correlation coefficient, r = 0.77, p 0.01) and latency (Pearson’s correlation coefficient, r = 0.59, p 0.05) of wave V responses to clicks between the left and right ear were well correlated in the control group. However, the amplitude (Pearson’s correlation coefficient, r = 0.05, p = 0.86) and latency (Pearson’s correlation coefficient, r = 0.66, p = 0.40) were not well correlated between ears in the tinnitus group. This lack of correlation may be indicative of tinnitus and/or a hidden hearing loss associated with tinnitus. Alternatively, it could be because of the limited data available from tinnitus participants.

The fact that a lower proportion of tinnitus participants elicited detectable responses (33%) than in the control group (62%) may be important, but this difference was not significant (Chi-squared test, $\chi$² = 1.7, p = 0.19). There was no significant difference in the amplitude or latency of wave V responses between the control and tinnitus groups when testing the left or right ears (Mann-Whitney U-test, p $>$ 0.05 for each test).

Statistical testing revealed no significant differences in the wave latency of either wave I, III or V, for either click, 2 kHz TB or 4 kHz stimuli, at any intensity, between control and tinnitus groups (Fig. 10 and Table 6). However, significant differences were found between the tinnitus cohort when stratified by CTI grade (Fig. 10 and Table 6). CTI 1 group (slight tinnitus) had an elevated mean wave latency for several wave types (I, III, and V) and intensities for click and 4 kHz TB stimuli, compared to other CTI groups (CTI2 and CTI3, mild and moderate tinnitus, respectively) (One-way ANOVAs and post-hoc Bonferroni multiple comparison analysis, with an adjusted p-value of p = 0.0083) (Fig. 10 and Table 6).