Early Datalogging Predicts Cochlear Implant Performance: Building a Recommendation for Daily Device Usage

Abstract

Objective:

To quantify the effect of datalogging on speech recognition scores and time-to-achievement for a ‘benchmark’ level of performance within the first year, and to provide a data-driven recommendation for minimum daily cochlear implant (CI) device usage to better guide patient counseling and future outcomes.

Study Design:

Retrospective cohort.

Setting:

Tertiary referral center.

Patients:

337 adult CI patients with datalogging and speech recognition outcome data who were implanted between August 2015 and August 2020.

Main Outcome Measures:

Processor datalogging, speech recognition scores, achievement of ‘benchmark speech recognition performance’ defined as 80% of the median score for speech recognition outcomes at our institution.

Results:

The 1-month datalogging measure correlated positively with word and sentences scores at one-, three-, six-, and twelve-months post-activation. Compared to age, sex, and preoperative performance, datalogging was the largest predictive factor of benchmark achievement on multivariate analysis. Each hour/day increase of device usage at 1-month resulted in a higher likelihood of achieving benchmark CNC and AzBio scores within the first year (OR 1.21, p <0.001) as well as earlier benchmark achievement. Receiver operator characteristic (ROC) curve analysis identified the optimal datalogging threshold at an average of 12 hours/day.

Conclusions:

Early CI device usage, as measured by 1-month datalogging, predicts benchmark speech recognition achievement in adults. Datalogging is an important predictor of CI performance within the first-year post-implantation. These data support recommended daily CI processor utilization of at least 12 hours/day to achieve optimal speech recognition performance for most patients.

INTRODUCTION

Cochlear implants (CI) offer a hearing rehabilitation option for patients who derive inadequate benefit from hearing aids and other assistive devices. Despite their overall effectiveness, there remains considerable variability in individual speech recognition outcomes. Many studies and meta-analyses have evaluated patient-related factors to predict speech recognition outcomes for patients with these devices.

Factors for post-lingually deafened adults including age at implantation, duration of hearing loss, gender, etiology, and pre-operative speech recognition scores, which each demonstrate statistically significant but low-to-negligible associations with ultimate CI performance (1,2). Other factors such as spiral ganglion cell count, top-down processing, and neural health are difficult if not impossible to alter prior to CI evaluation or implantation (3–5). Surgical variables such as electrode position, electrode type, and surgical technique are implicated in outcomes but are imperfectly modifiable and only accessible in the immediate perioperative period (6–13). Post-operative elements include utilization of residual low-frequency acoustic hearing, advanced device mapping/programming, and the utilization of aural rehabilitation (14,15). Despite a multitude of potential factors contributing to speech performance, our ability to affect or improve CI performance is limited by individual patient characteristics, the limited windows of intervention for certain variables, as well as the cost associated with therapies, imaging, and clinical encounters.

Recent improvements in CI processor features allow for the monitoring of average daily processor usage or “datalogging” within the programming software. Initially, clinicians utilized this feature to monitor wear time for pediatric patients, demonstrating that average daily processor usage had a significant association with auditory outcome measures in children (16–18). Alhabib et al. suggested that daily usage greater than 8.3 hours per day was associated with acceptable language development in this population (19). In the adult population, correlation coefficients of daily processor use with speech recognition are between 0.4–0.6 (15,20). A correlation of 0.6 is stronger than associations of other commonly referenced factors such as age at implantation (21,22), duration of deafness (21–24), electrode position (25–27), and is generally equivalent to correlations of spectral resolution (28–30) and spectrotemporal resolution (31,32). Recent research demonstrates that intervention to increase daily CI usage results in improved auditory perception and speech recognition scores for adult CI users even after years of device usage, suggesting a causal relationship between daily CI use and auditory outcomes (33).

Despite a relatively recent appreciation for daily device usage as a contributor to auditory perception and speech recognition scores in adults, a paucity of evidence exists regarding the impact of CI device datalogging on the ultimate success and speech outcomes in adults. Prior recommendations advising at least 10 hours of daily device usage are based on better-than-average speech recognition outcomes for adults above the 50^th-percentile for datalogging. However, little is known regarding the importance of early (e.g., the first month immediately after activation) usage or any additional benefit to datalogging beyond 10 hours per day. We hypothesized that early datalogging would predict CI performance in adults and that additional benefits to processor daily usage extend beyond 10 hours per day. Herein, we aim 1) to quantify the effect of datalogging on speech recognition scores and time-to-achievement for a ‘benchmark’ level of performance within the first year, and 2) to provide a data-driven recommendation for minimum daily CI device usage to better guide patient counseling and future outcomes.

MATERIALS AND METHODS

After institutional review board approval, a retrospective review identified adult CI recipients implanted with their first device between August 2015 and August 2020 with datalogging information available at the 1-month post-activation visit. Patient demographics, datalogging, and speech recognition outcomes were collected at follow-up visits within the first-year post-activation (most commonly 1-, 3-, 6-, and 12-months). Datalogging was defined as the average hours of processor use per day regardless of listening condition; this choice was motivated by data suggesting a much lower correlation between speech recognition versus time in speech than speech recognition versus total datalogging (34). Exclusion criteria included those with prelingual deafness, prior ipsilateral or contralateral CI surgery, device failure during the study period, or absence of datalogging data (often due to lack of processor feature).

Adult patients underwent audiologic evaluation according to the Minimum Speech Test Battery (35). Speech recognition testing for Consonant-Nucleus-Consonant (CNC) words and AzBio sentences in quiet was completed in a sound booth with speech presentation level of 60 dB sound pressure level (SPL, A-weighted) (36,37). Stimuli were presented from a single loudspeaker at 0-degree azimuth (1 m). Testing was completed in the CI-alone condition with the contralateral ear plugged or masked as required based on the thresholds measured for the contralateral ear.

The calculation of time-to-benchmark for speech recognition scores in this cohort was based on a prior report suggesting patients achieved benchmark at or above 80% of the average speech recognition scores for patients undergoing CI (38). This time-to-benchmark metric was designed to create a binary classification system for CI achievement. During this retrospective chart review, we captured data for all post-lingually deafened adults undergoing implantation at our institution from 2015-2020 who also had speech recognition scores available at 12-months post-activation; the mean CNC and AzBio in quiet scores at 12 months were 50.3% and 64.7%, respectively (n=405). As a result, benchmark CNC and AzBio in quiet scores were calculated as a CNC score of 40% or greater and AzBio in quiet score of 52% or greater.

Analyses were performed using IBM SPSS Statistics version 28.0 (IBM Corp., Armonk, New York) and GraphPad Prism 9.2.0 (GraphPad Software, La Jolla, California). Nominal data were summarized using frequency distributions. Given the presence of some moderate outliers and skewness of the age, CNC, AzBio, datalogging, and time to benchmark data distributions, median, inter-quartile range, and range were used for summarizing distributions. Unadjusted (bivariate) and adjusted (multivariate) logistic regression models were used to test for associations of age, sex, preoperative CNC/AzBio scores, and post-activation CI achievement of benchmark scores within the first one year. Receiver operating characteristic (ROC) models were used to generate initial CI usage thresholds likely to be predictive of achieving those respective. CNC or AzBio benchmarks. Estimates of expected time to benchmark and differences between groups were generated using Kaplan-Meier analyses and log-rank Mantel-Cox tests. Patients who did not meet either benchmark within one year were censored from these analyses. P values <0.05 were considered statistically significant.

RESULTS

Demographics and Clinical Characteristics

Characteristics of the 337 patients with 1-month datalogging data and at least one post-activation CNC or AzBio score are summarized in Table 1. Median age was 68.7 years and 53% were male. The median pre-operative CNC word score was 8% (IQR 0% to 22%) and the median AzBio in quiet sentence score was 10% (IQR 0% to 34%). Median 1-month CI use for all patients was 10.7 hours per day, with a range of 0.2 to 17.3 hours/day.

Table 1.

Patient Characteristics and Demographics

Characteristic	Incidence/Median (n=337)	Range
Gender
Female	157 (47%)
Male	180 (53%)
Laterality
Left	146 (43%)
Right	191 (57%)
Age	68.7	(IQR 56-78, range 17.5-96.7)
Pre-operative CNC word score	8.0	(IQR 0-22, range 0-71)
Pre-operative AzBio in Quiet sentence (n=292)	10.0	(IQR 0-34, range 0-80)
1-month CI use (hours per day)	10.7	(IQR 8.0-12.6, range 0.2-17.3)
Months to CNC Benchmark* (n = 238, 70.6%)	2.1	(IQR 1.0-3.7, range 1-12)
Months to AzBio Benchmark* (n = 209, 71.6%)	3.4	(IQR 3.0-5.3, range 1-12)

CNC indicates Consonant-Nucleus-Consonant word score; IQR, interquartile range.

Early Datalogging Correlates with Datalogging and Speech Outcomes Within the First Year

One-month datalogging values correlated to datalogging values at 3-months (r_s=0.7646, p <0.0001), 6-months (r_s=0.6996, p <0.0001), and 12-months (r_s=0.5300, p <0.0001) (Figure 1). One-month datalogging values significantly correlated to speech outcomes at the 1-month, 3-month, 6-month, and 12-month follow-up visits (Figure 2).

Figure 1.

Datalogging at 1 month plotted against datalogging measured at subsequent follow-up visits. The bold line marks equivalence. The 3-month (r_s=0.7646, p <0.0001), 6-month (r_s=0.6996, p <0.0001), and 12-month (r_s=0.5300, p <0.0001) points are shown.

Figure 2.

Daily usage at 1 month is positively correlated with A) 1-month, B) 3-month, C) 6-month, and D) 12-month consonant-nucleus-consonant (CNC) word scores as well as E) 1-month, F) 3-month, G) 6-month, and H) 12-month AzBio Quiet sentence scores. Spearman correlation coefficients, p values, and number of data points (n) are given.

Achievement of Benchmark Scores and Associated Factors

Within 1 year of activation, 70.6% (238 of 337) of patients met the 40% CNC performance benchmark and 71.6% (209 of 292) of patients met the 52% AzBio in quiet performance benchmark. Factors associated with benchmark attainment are included in Table 2. Summaries of the unadjusted univariate logistic regression analyses, as well as multivariate models that included age, sex, pre-operative score, and 1-month datalogging are also shown in Table 2. Higher pre-operative word recognition scores increased the likelihood of achieving the respective benchmarks in both the univariate and multivariate analyses. Relative to younger patients, older patients were less likely to meet the AzBio in quiet benchmark but not the CNC benchmark within the first year. Based on the calculated odds ratios, every hourly increase in datalogging increased the likelihood of meeting benchmark by approximately 1.2-fold (p < .001). Patients who met the CNC benchmark within the first year utilized their CI a median of 11.3 hours per day (IQR = 9.1-12.9) compared to a median 9.1 hours (IQR = 5.2 - 11.3) for those who did not meet that benchmark. Use patterns were similar for the AzBio benchmark (median = 11.3 vs. 9.5 hours/day). Within the first year 88 patients did not meet either benchmark, 30 met the CNC but not the AzBio benchmark, and 10 met the AzBio but not the CNC benchmark.

Table 2.

Associations with Benchmark Achievement

		Unadjusted		Adjusted
		OR	P value	OR (95% CI)	P value
CNC (N=337)	Age (years)	0.99	.584	0.99 (0.97, 1.01)	.480
	Male gender	1.49	.100	1.16 (0.69, 1.93)	.566
	Word Score (Pre-operative)	1.03	.001	1.03 (1.01, 1.06)	.001
	Datalogging (1-month)	1.20	< .001	1.21 (1.12, 1.31)	< .001
AzBio in Quiet (N=292)	Age (years)	0.97	.003	0.97 (0.95, 0.99)	.002
	Male gender	0.93	.779	0.77 (0.43, 1.35)	.354
	Sentence Score (Pre-operative)	1.02	.007	1.02 (1.01, 1.04)	.006
	Datalogging (1-month)	1.19	< .001	1.21 (1.11, 1.31)	< .001

CI indicates confidence interval; CNC, Consonant-Nucleus-Consonant word score; OR, odds ratio.

Building Statistical Models for Daily Device Usage Recommendations

ROC curve analysis was performed to build classification models for both CNC [area under the curve (AUC) = 0.67, 95% C.I.: 0.60 0.73, p < .001] and AzBio in quiet (AUC: 0.67, 95% C.I.: 0.60, 0.74, p < .001) benchmarks (Figure 3). Consistent with the logistic regression analyses, increased datalogging within the first month of usage was associated with an increased likelihood of meeting both the CNC and the AzBio benchmarks. Utilizing our group’s prior published work, we chose an initial cut-off of ≥10 hours per day as a starting point for analysis (20). Evaluation of the sensitivity-specificity curves suggested usage of ≥10 hours provided an initial indicator of increased likelihood of achieving the CNC or the AzBio benchmark within 1 year of CI activation. Subsequent multivariate Cox regression models indicated a significantly shorter time to achieving benchmark for those patients who used the CI ≥10 hours/day within the first month after activation (p < .001). Further analysis of the ROC curves was performed utilizing the Positive Likelihood Ratio (LR+), defined as Sensitivity / (1 – Specificity) (39). In this clinical situation, LR+ is the probability that a patient would attain benchmark speech recognition scores within the first year. As a result, our strategy to determine an ideal datalogging threshold was to identify the point along the curve that resulted in the highest LR+ at the lowest relative hour/day datalogging. Via this approach, we identified the optimal datalogging thresholds at 11.5 hours/day for CNC words and 12.5 hours/day for AzBio sentences in quiet.

Figure 3.

Receiver operating characteristic (ROC) curve analysis was performed to build classification models for both A) Benchmark achievement of CNC word score [area under the curve (AUC) = 0.67, 95% C.I.: 0.60 0.73, p < .001] and B) Benchmark achievement of AzBio Quiet sentence score (AUC: 0.67, 95% C.I.: 0.60, 0.74, p < .001). Sensitivity is shown on the y-axis, 1-specificity on the x-axis, and hour per day thresholds are marked on the corresponding points on the curve.

Based on prior recommendations and the interpretation of the ROC curves, time-to-event curves for CNC benchmark attainment were designed based on the 10 hours/day threshold (Figure 4A) and the 12 hours/day threshold (Figure 4B). The time-to-event curves were significantly different for both thresholds based on log-rank (Mantel-Cox) tests (p < 0.0001). Hazard Ratios (HR) were 1.883 (95% C.I. = 1.472, 2.408) for the 10 hours/day threshold and 3.189 (95% C.I. = 2.307, 4.409) for the 12 hours/day threshold. Ultimately, 84% (96/114) of patients with 1-month datalogging ≥12 hours/day met benchmark within the first year as compared to 80% (166/207) with 1-month datalogging ≥10 hours/day.

Figure 4.

Time-to-event curves for CNC benchmark attainment were designed based on the A) 10 hours/day threshold and the B) 12 hours/day threshold. The time-to-event curves were significantly different for both thresholds based on log-rank (Mantel-Cox) tests (p < 0.0001). Hazard Ratios (HR) were 1.883 (95% C.I. = 1.472, 2.408) for the 10 hours/day threshold and 3.189 (95% C.I. = 2.307, 4.409) for the 12 hours/day threshold.

Based on datalogging, the median time to CNC benchmark was: <10 hours/day: median time = 6.6 months (95% C.I. = 3.0, 10.1), 10-11.9 hours/day: median time = 3.1 months (95% C.I. = 2.6, 3.7), ≥12 hours/day: median time = 2.3 months (95% C.I. = 0.9, 3.6). Compared to the group with datalogging 10-11.9 hours/day, those patients utilizing their devices ≥12 hours/day had a significantly shorter median time to benchmark (unadjusted: p = .034, adjusted for age, sex, pre-operative CNC: p < .001).

With regards to AzBio sentences in quiet, the median time to benchmark score was calculated for <10 hours/day: median time = 6.8 months (95% C.I. = 3.9, 9.6), 10-11.9 hours/day: median time = 3.5 months (95% C.I. = 2.9, 4.1), ≥12 hours/day: median time = 3.6 months (95% C.I. = 3.3, 3.9). Compared to the group with datalogging 10-11.9 hours/day, those patients utilizing their devices ≥12 hours/day had a significantly shorter median time to benchmark for AzBio sentences in quiet as compared when adjusted for age, sex, and pre-operative AzBio quiet score (p = 0.012) but not for the unadjusted calculation (p = 0.141).

DISCUSSION

Recent investigations have focused on characterizing the role of discrete patient factors, device properties, and surgical techniques as they relate to speech recognition outcomes for patients with CIs. Unfortunately, many of the patient-related factors such as age, duration of deafness, and neural health are not readily amendable at time of CI evaluation. Similarly, considerations such as electrode type, scalar location, and preservation of residual cochlear function are difficult to modify after CI surgery. Post-implantation interventions are limited but include CI mapping, consistent daily processor use, fitting of electroacoustic stimulation , and aural rehabilitation.

While obtaining adequate CI experience through consistent wear time may seem intuitive, a paucity of data prevents strong, evidence-based recommendations for the optimal level of daily processor wear time for adult CI recipients. Clinicians often recommend that patients wear their CI processors during “all waking hours,” but a data-driven numerical recommendation may enable more tangible wear time goals—similar to using a fitness tracker to motivate specific step goals or exercise minutes per day. Prior work from our group has suggested that greater than 10 hours of CI use per day was associated with better than average speech recognition outcomes; we sought to further interrogate this relationship through early processor utilization data and speech recognition outcomes to form a more robust clinical recommendation.

Unsurprisingly, 1-month datalogging correlated with datalogging recorded at the 3-, 6-, and 12-month clinic visits. We theorize that daily use habits are established early and persist through at least the first year. Higher early datalogging correlated with greater speech recognition performance at each timepoint, with the correlation coefficients and slopes for the best-fit lines at each time of assessment not significantly different. Together, these data suggest that early device datalogging predicts both for future processor utilization and speech recognition outcomes.

To formulate a daily processor usage recommendation, establishment of a binary classification system through benchmark achievement is necessary to perform ROC curve analysis. This strategy places less importance on raw performance and more emphasis on benchmark attainment to ascertain, ‘Did the patient attain reasonable speech recognition success with their CI within the first year?’

We initially chose a recommendation of at least 10 hours/day CI device usage based solely on prior work suggesting that patients above this threshold perform better than average (20). However, to answer ‘What specific hours/day device usage do we recommend for patients to get the best speech recognition outcomes after CI surgery?’ we must combine the more statistical model afforded by ROC curve analysis with appropriate clinical assumptions. For the average patient, we assume that daily CI processor utilization is not incredibly difficult, burdensome, or expensive to influence for a reasonable portion of the 24-hour period (e.g., 12 hours/day versus 10 hours/day). ROC curve analysis demonstrates the highest LR+ for meeting benchmark score with the lowest hours/day usage at approximately 11.5 hours/day for CNC and 12.5 hours/day for AzBio in quiet, or about 12 hours/day overall. This practical utilization of the ROC curve, combined with our raw speech recognition data, suggests a relative optimization for benchmark achievement of speech recognition metrics for patients with >12 hours/day datalogging at the 1-month time point.

In 2017, Busch et al. analyzed datalogging for 1,501 adult and pediatric CI recipients at multiple timepoints after implantation and found daily average wear time to be 10.7 hours, with a 95% confidence interval of 4.3 to 15.2 hours/day (40). This mean processor utilization and variability in wear time is remarkably similar to our cohort’s 1-month data. Recently, a questionnaire was designed and validated for the assessment of factors affecting daily CI usage (41). Clinicians may consider using this questionnaire at the 1-month post-activation appointment to explore potential barriers to CI use. Further inquiries regarding the large variability in wear time as well as the factors affecting early CI usage are indicated.

A separate consideration is the possibility of a ‘learning phase’ for post-lingually deafened adults suggested to primarily occur within the first 6 months post-implantation, after which CI recipients may enter a plateau period of relative stability in speech recognition scores.(42) This is in contrast to research demonstrating that patients utilizing their device less than 10 hours/day had significantly improved speech recognition scores after just 4 weeks of increased datalogging, even years post-implantation.(33) Despite our evolving understanding of the temporal requirements for neuroplasticity and CI adaptation after implantation, the early post-operative period is a great opportunity for clinician intervention and counseling regarding daily device usage. Further research is necessary to elucidate the presence of any critical period for daily CI device usage as well as the relationship between datalogging and speech recognition outcomes, especially for patients lacking full-time usage who wish to improve their performance.

This study was limited by its retrospective nature. We did not separate patients based on duration of deafness, etiology, or other factors that could potentially influence their ultimate speech recognition outcomes. We acknowledge certain assumptions related to our statistical methods, although we have also included their rationale. While unfamiliar, the notion of a time-to-benchmark datapoint for CI recipients is not novel and allows ROC analysis. Further research is needed to investigate the effects of datalogging on raw speech recognition scores and performance. Given that the work herein captured daily device use without consideration of the user’s listening environment, further investigation is warranted with regards to the specific environments of CI utilization.

CONCLUSION

Herein, multivariate logistic regression analyses demonstrate a strong association between early CI processor usage datalogging and the likelihood for attainment of benchmark scores for CNC and AzBio in quiet speech recognition tests within the first-year post-implantation. Patients with higher 1-month datalogging also reach these benchmarks earlier in the post-implantation period. ROC curve analysis suggests the highest likelihood ratio for benchmark score attainment is above 12 hours/day. Patient counseling to increase CI processor datalogging may be an accessible, low-cost, and low-burden intervention for improving outcomes. In the absence of other factors that might limit a patient’s ability to wear their device, we recommend daily CI processor utilization of at least 12 hours/day (an increase beyond the previously published recommendation of 10 hours/day). Further work is necessary to examine the raw scores, long-term speech outcomes, time to performance plateau, and other factors affecting variability in patient datalogging.