a standardized CI fitting method individualizing programming in an efficient ... programming interface; this was made available to us .... Custom Sound) so that.
Speech Based Optimization of Cochlear Implants Alice E. Holmes* Lee Krause***Rahul Shrivastav** Hannah W. Siburt* *Department of Communicative Disorders, College of Public Health & Health Professions, Gainesville, Florida; **Communication Sciences and Disorders, College of Liberal Arts & Sciences, Gainesville, FL; *** Audigence, Inc. The optimization routine resulted in a overall reduction in the net weighted DF error and improvement in all the outcome measures for both Opt I and Opt 2. Individual outcome performance is shown in Figures 2, 3 and 4.
SESSION 1 (Cont) 4. The optimization routine (Figure 1) was run for a period of 30 minutes with the following:
We have developed a novel fitting approach utilizing optimization theory to rapidly “scan” a patient’s performance at multiple device settings, thereby creating a multidimensional “performance space” for each individual patient. The optimization routine is designed to find settings that minimize speech recognition errors using real-time Distinctive Feature (DF) analysis (Jakobson, et al., 1963) of phoneme errors. Unlike conventional approaches to fitting, the device is adjusted based on phoneme recognition abilities and not subjective quality. This approach may provide a standardized CI fitting method individualizing programming in an efficient manner.
90
90
80
80
70
AI Algorithm
70
Optimization Models Rate
Learning feedback
Gain Baseline Best Outcome
50
40
30
Gain Baseline Best Outcome
50
40
20
CI Loudness Growth
20
10
FAT 10
0
12.50 -10
0
Subjects Ordered by Baseline Performance
c) The next combination of FAT, PR & LG was automatically recommended and tested.
60
Subjects Ordered by Baseline Performance
Figure 2. Individual CNC-Phoneme performance. Figure 1. Optimization Schematic.
Figure 3. Individual CNC-Word performance.
BKB-SIN (dB) 20
d) This procedure was repeated until the end of the 30 minute optimization session 5. The combination of LG, PR and FAT resulting in the smallest NWE was selected as Optimization 1 (Opt 1) and programmed into their speech processor.
2.50
Gain Baseline Best Outcome
5
0.00 BKB Baseline
*Both Opt 1 and baseline maps were saved on the CI device which participants were free to use.
-10
Figure 4. Individual BKB-SIN performance.
Statistical Analysis SESSION 2 (~2 weeks after Session 1)
The experiment was conducted as a repeated measures design where each subject was compared to their own baseline. Each subject was tested in 3 sessions, approximately two weeks apart. All test procedures were conducted following UF IRB approval.
SESSION 3 (~2 weeks after Session 2) Age (Years) Mean S.D. Range Length of CI Use (months) Mean S.D. Range Type of CI N24 New Freedom
57.3 19.9 24-82
The ANOVA-R indicated a significant difference among the three conditions using both a Sphericity Assumed analysis (p < 0.003) and the more conservative GreenhouseGeisser analysis (p < 0.004). Further trend analyses indicated a significant ascending omnibus trend from baseline (p < 0.004) and a non-significant ascending quadratic trend from baseline (p > 0.05). Pairwise comparisons showed significant differences between baseline and Opt 1 (p < 0.025) and between Baseline and Opt 2 (p < 0.015).
1. Outcome performance using Opt 2 was evaluated using CNC lists in quiet and BKB-SIN measurements as reported above. 2. Subjects then chose the maps that they wanted saved in their speech processors for regular/everyday use.
25.6 28.97 5-115
RESULTS
60.00
1. The T- and C- values were obtained at multiple pulse rates. This was necessary before the optimization routine manipulated pulse rates (PR). 2. Baseline performance using the subject’s current map on open-set recognition was evaluated using CNC lists in quiet and BKB-SIN measurements. All stimuli were presented through a direct connection to the speech processor and at a constant level across all test sessions (approximately 60 dBA). 3. The Optimization program was designed to interface with a customized version of the manufacturer’s programming software (Cochlear Corp. Custom Sound) so that programming changes recommended by the algorithm could be tested seamlessly.
0.00 CNC Word RAU Baseline
CNC Word RAU Opt 1 Session 2 CNC Word RAU Opt 2 Session 3
Error bars: +/- 2 SE
Subject Map Parameters (Number of Subjects)
CNC Phoneme ANOVA-R
Table 1: Subject demographics
SESSION 1
Figure 5. Means and standard error for Rau transformed CNC-word Scores.
40.00
20.00
Table 2 shows a summary of the subjects baseline and optimized maps. N=3 N=17
Stimulation Rate 500 720 900 1200 1800 5 3 8 2 1 4 8 4 1 1 2 5 4 2 5 Loudness Growth (LG) 10 15 20 25 30 Baseline 1 18 1 Opt 1 6 2 5 5 2 Opt 2 4 3 3 5 5 Frequency Allocation Table (FAT) 188-7938 188-7438 188-6938 188-6563 188-6063 188-5938 188-5813 Baseline 19 1 Opt 1 6 1 8 1 2 2 Opt 2 4 6 4 2 1 1 2 Baseline Opt 1 Opt 2
250 1 2 2
At the end of this clinical trial, 17 out of 20 patients preferred to continue using one of their optimized maps. Subjective ratings were also obtained form each subject and are currently being tabulated.
CONCLUSIONS
CNC Word ANOVA-R
Mean
Procedures:
*Both Opt 1 and the subject’s choice of either baseline or Opt 1maps were saved on the CI device which participants were free to use.
A separate Analysis of Variance with Repeated Measures (ANOVA-R) was completed for each of the outcome measures using the following model: 20 subjects in 3 conditions (Baseline; Optimization 1 after 2 weeks experience; Optimization 2 after 2 weeks of experience). All percent scores (CNC word and phoneme) were transformed into Arcsine units (RAU Transform) to stabilize the error (Studebaker, 1985).
The CNC Phoneme ANOVA-R indicated a significant difference among the three conditions using both a Sphericity Assumed analysis (p < 0.003) and the more conservative Greenhouse-Geisser analysis (p < 0.008). Further trend analyses indicated a significant ascending omnibus trend from baseline (p < 0.015) and a significant ascending quadratic trend from baseline (p < 0.035). Pairwise comparisons showed significant differences between base line and Opt1 (p < 0.003) and between Baseline and Opt 2 (p < 0.04).
The optimization method used in this study resulted in improved subject performance in all outcome measures. Speech perception was significantly better in word and phoneme identification with optimized maps. In addition, subjects performed better in noise using the optimized maps. All subjects entered the study wearing CIs programmed in the traditional clinical approach. However, as seen in Table 2, significant individual variability was found between baseline and optimized map parameters. The current clinical method for cochlear implant programming is both subjective and time consuming. For example, in the current study up to 40 maps were tested in a 30 minute optimization session. In contrast, testing the same number of maps using the traditional method would have taken multiple programming sessions with the clinician relying on patient report rather than objective data. The optimization process described here was successful in greatly reducing the time necessary to optimally program a CI with improved patient performance. In addition to obtaining CNC and BKB-SIN test outcome measures, subjective measures were also gathered from our subjects. Subjective feedback regarding the optimization procedure and its results were both extremely positive. Feedback included the report of increased clarity, increased ability to talk on the telephone and watch television, as well as better understanding in the presence of background noise. Subjective questionnaires with situation specific data are currently being analyzed statistically. Future applications using the optimization method include expanding the number of parameters adjusted during the procedure, and increasing the CI population eligible for optimization. In addition, this method is being investigated for use with other hearing devices such as hearing aids.
100.00
ACKNOWLEDGEMENTS
80.00
Figure 6. Means and standard error for Rau transformed CNCPhoneme Scores.
60.00
Mean
20 experienced CI users were recruited from the University of Florida Speech & Hearing Clinic. All participants used a Freedom implant processor (Cochlear, Inc.). This was essential because the optimization routine needs access to the device programming interface; this was made available to us by Cochlear, Inc. Participants received monetary compensation for their time.
BKB Opt 2 Session 3
-5
2. The Optimization procedure described above was then repeated to obtain Optimization 2 (Opt 2) and programmed into their speech processor.
Subject Demographics Gender Male N=7 Female N=13
BKB Opt 1 Session 2
Error bars: +/- 2 SE
6. Subjects were asked to use the optimized map until Session 2.
3. Subjects were asked to use the optimized map until Session 2. Participants:
Figure 7. Means and standard error for BKBSIN signal-to-noise ratios.
7.50
10
0
1. Outcome performance using Opt 1 was evaluated using CNC lists in quiet and BKB-SIN measurements as reported in Session 1.
METHODS
10.00
5.00
15
Subjects Ordered by Baseline Performance
In the present study, we report findings of our initial clinical trial. At this time, the optimization algorithm is designed to optimize the following parameters [pulse rate (PR), loudness growth (LG) & frequency allocation table (FAT)]. However, the optimization routine is being modified to include additional parameters.
The BKB-SIN ANOVA-R indicated a significant difference among the three conditions using both a Sphericity Assumed analysis (p < 0.02) and the more conservative Greenhouse-Geisser analysis (p < 0.03). Further trend analyses indicated a non-significant ascending omnibus trend from baseline (p > 0.05) and a significant ascending quadratic trend from baseline (p < 0.009). Pairwise comparisons showed significant differences between base line and Opt 1 (p < 0.03) and a non-significant difference between Baseline and Opt 2 (p = 0.195).
30
Inputs from
software
60
Performance/Gain
b) A unique optimization metric, summarized as the “net weighted feature error” (NWE) for the processor setting was calculated by the optimization algorithm.
100
BKB-SIN ANOVA-R
Mean
Most audiologists fit CI speech processors to manufacturer recommended settings. These can be modified based on a patient’s feedback regarding perceived overall sound quality and not actual performance on speech intelligibility tests.
a) A series of VCV syllables were presented and the subject’s verbal responses were recorded by the researcher.
CNC-Word in Quiet (%)
100
Performance/Gain
Modern CI devices allow multiple programmable parameters. In a number of studies that have manipulated parameters such as pulse rate (PR), number of electrodes used, and pulse width (PW), no single parameter setting has been shown to be universally optimal for all patients (Loizou, et al., 2000; Holden, et al., 2002; Skinner, 2003). This makes it difficult for audiologists to know which parameters to focus on during implant mapping.
CNC-Phoneme in Quiet (%) DF Error Matrix
Performance/Gain
INTRODUCTION
40.00
This project is funded by Audigence, Inc. and the Florida High Tech Corridor Council. We wish to thank Cochlear Corporation for supplying the fitting software platform and for their extensive and timely technical support.
20.00
0.00 CNC Phoneme RAU Baseline
CNC Phoneme RAU Opt 1 Session 2
Error bars: +/- 2 SE
CNC Phoneme RAU Opt 2 Session 3
We also want to thank our subjects for their willingness to participate in the experiment.
