Source-linked AI summary
Estimates of neural health using cochlear implant digital twins correlate with speech recognition
Erin L. Bratu, Andrea J. DeFreese, Katelyn A. Berg, Margaret Wilson, Aaron C. Moberly, Robert F. Labadie, Jack H. Noble
TL;DR
Cochlear implant outcomes remain highly variable, while limited and overlapping stimulation constrains the electro-neural interface. The paper builds patient-specific digital twins and finds that estimated neural health explains substantial speech-recognition variability; preliminary remapping results also report improved scores.
Problem
CI outcomes remain highly variable, and some recipients receive little benefit beyond basic sound awareness and lip-reading assistance.
Source: §I · p. 1Method
The paper develops patient-specific digital twins that model spatial, electrical, and neural-health properties of the cochlear implant interface.
Source: §IV · p. 7Results
Neural-health scores explain 50–60% of speech-recognition variability in quiet and noisy conditions in this dataset.
Source: §IV · p. 8Takeaways & Limitations
Digital twins could be useful for programming pediatric recipients or others who have difficulty providing subjective feedback.
Source: §IV · p. 8The compound-action-potential model assumes synchronous additive responses, likely underestimating healthy-fiber counts, and models only steady-state neural activation.
Source: §IV · p. 8
Abstract
from arXiv · showhide
The cochlear implant is a neural prosthetic that uses an array of electrodes to directly stimulate auditory nerve fibers with electrical impulses. One factor that is known to impact outcomes is the quality of the electro-neural interface. In this work, we propose a pipeline that permits creating accurate and patient-specific digital twins that model the cochlear implant electro-neural interface. We leverage medical images and electrophysiological measurements from the patient's device for digital twin construction. We validate our digital twins by comparing predictions of neural response measurements to clinical measurements and speech recognition outcomes. Neural health estimated by the digital twin correlates strongly with word recognition rates (R=0.75, p<1e-5) in an analysis of 29 cases and sentence in noise recognition rates (R=0.78, p<1e-4) in an analysis of 21 cases. We also report preliminary findings with digital twin-based patient remapping with 6 subjects, where word recognition rates improved on average by 10 percentage points. These digital twins have the potential to provide clinicians with an unprecedented window into the electro-neural interface for individual cochlear implant recipients, which could lead to improved hearing outcomes when used as a foundation for program customization.
I. INTRODUCTION
CI outcomes remain highly variable, and limited, overlapping electrode stimulation constrains the electro-neural interface. The paper proposes patient-specific digital twins as a way to characterize that interface and inform customized programming.
I. INTRODUCTION
- Despite technological advances, CI outcomes remain highly variable, and some recipients receive little benefit beyond sound awareness and lip-reading assistance.Source: §I · p. 1
- Electro-neural interface quality is known to affect CI outcomes, and audiologists can adjust a recipient’s electrode set and stimulus parameters through the patient’s map.Source: §I · p. 1
- Traditional map creation relies primarily on subjective recipient feedback, and adapting to a new map may take weeks or months.Source: §I · pp. 1–2
- An individual interface depends on anatomy, electrode position, tissue electrical properties, nerve health, and time-dependent neural behavior, including refractoriness lasting 1–3 ms.Source: §I · p. 2
- At most 22 contacts must serve up to around 30,000 auditory nerve fibers, while broad, overlapping stimulation can smear spectral information and produce channel interactions [11, 12].Natural hearing provides a higher degree of frequency selectivity.Source: §I · p. 2
- The paper proposes algorithms for patient-specific digital twins of the electro-neural interface, which could give clinicians a view of that interface to support customized programming.Source: §I · p. 2
A. Previous Work
Prior cochlear models combine patient-specific spatial representations, electrical field simulations, and neural activation models, but approaches vary in their customization and scale.
A. Previous Work
- Prior models often adapt detailed cochlear anatomy from cadaver imaging to lower-resolution in-vivo scans, while electrode positions can be localized from postoperative CT.Source: A. Previous Work · p. 2
- Volume conduction has been modeled with finite element, boundary element, and finite difference methods, usually using fixed tissue resistivity despite the possibility of patient-specific electrical customization.Source: A. Previous Work · p. 2
- Single-nerve activation models include Rattay and generalized Schwarz-Eikhof-Frijns models, but simulating up to 30,000 fibers individually quickly becomes impractical.Source: A. Previous Work · p. 2
- Bundle-level phenomenological models summarize responses across hundreds or thousands of fibers, while prior groups had not explored patient-specific neural health representation at that level.Source: A. Previous Work · p. 2
II. METHODS
The paper proposes a patient-customized digital-twin pipeline that combines patient information to model the spatial, electrical, and neural-health properties of the electro-neural interface.
II. METHODS
- The proposed pipeline customizes digital twins to individual patients, capturing spatial, electrical, and neural-health properties of the interface.Source: §II · p. 2
- Digital-twin construction leverages medical images and electrophysiological measurements from the patient.Source: §II · p. 2
- The authors validate the twins by comparing predicted neural-response measurements with clinical measurements and outcomes.Source: §II · p. 2
A. Data
The study collected imaging and electrophysiological data from cochlear implant recipients to construct and characterize their digital twins.
A. Data
- For each of 30 recipients, the study collected pre- and post-implantation CT to localize ear anatomy and electrode contacts.Existing validated methods were used for anatomical localization, electrode localization, and image registration [18, 19, 30-36].Source: A. Data · p. 2
- Stimulation Current Induced Non-Stimulating Electrode Voltage measurements record voltages at non-stimulating electrodes when one electrode is activated as a current source.Source: A. Data · p. 2
- Neural response telemetry included amplitude growth functions on every contact and spread-of-excitation functions on even-numbered electrodes.Source: A. Data · p. 2
2. CI electric field model
The digital twin pipeline has spatial, electrical, and ANF model stages. AGF measurements estimate neural health and thresholds, while SOE measurements test predictions and reflect stimulation overlap between contacts.
2. CI electric field model
- The pipeline comprises three stages: spatial model, electrical model, and ANF model.Source: Figure 1 · p. 3
- SOE uses one electrode for stimulation and other electrodes as maskers, measuring shared ANF stimulation overlap between contact pairs.Source: §2 · p. 3
- AGF measurements estimate patient-specific neural health and thresholds, while SOE measurements are held out to test model predictions.Source: §2 · p. 3
B. Digital Twin Creation
The digital twin combines patient-specific cochlear geometry and electrical properties with neural-response models. Clinical electrophysiological measurements calibrate neural-health estimates and test predicted responses.
B. Digital Twin Creation
- The pipeline comprises spatial, electrical, and ANF model stages.Source: B. Digital Twin Creation · p. 3
- A statistical shape model segments cochlear structures from preoperative CT, while postoperative CT supports electrode-array segmentation and localization.The spatial model also constructs ANF paths from prior knowledge of fiber morphology because the fibers cannot be visually resolved.Source: B. Digital Twin Creation · pp. 3–4
- The electrical model builds tissue-class maps from cochlear segmentations and solves for voltage; patient-specific resistivities are tuned against clinical SCINSEV measurements.Source: B. Digital Twin Creation · p. 4
- A gSEF biophysical model simulates single-fiber action potentials from voltage maps sampled along modeled ANFs.The resulting membrane currents are scaled by transfer resistances to estimate the recorded single-fiber action potential.Source: B. Digital Twin Creation · pp. 4–5
- Single-fiber thresholds are adjusted using AGF thresholds, with four smoothly varying RBF control points scaling predictions across bundles.The resulting threshold map retains the gSEF-predicted structure while matching patient-specific threshold behavior more closely.Source: B. Digital Twin Creation · p. 5
- A cumulative distribution estimates the activated proportion of each bundle, and bundle action potentials are summed to predict compound responses.Source: B. Digital Twin Creation · p. 5
- For steady-state AGF and SOE measurements, available fibers are represented by H_f, the estimated healthy-fiber count; temporal neural behavior is left unmodeled.Source: B. Digital Twin Creation · p. 5
- Healthy-fiber estimates are fitted to clinical AGF data with weighted squared error and regularized for smooth variation across bundles.The objective weights above-noise-floor eCAPs more heavily; health parameters use 12 RBF control points.Source: B. Digital Twin Creation · p. 5
C. Experiments
Experiments assess held-out SOE prediction, associations between estimated neural health and speech recognition, and preliminary remapping in six participants.
C. Experiments
- 30 recipients had the data required to create digital twins, and 29 also had SOE measurements for evaluating predictions held out from training.Source: C. Experiments · p. 6
- The speech analysis relates neural health scores to CNC word recognition in quiet and AzBio sentence recognition at +5 dB SNR in multi-talker babble.Included participants had at least six months of device use.Source: C. Experiments · pp. 6–7, Figure 4 · p. 7
- Figure 3 summarizes predicted-versus-clinical SOE eCAP RMS errors across 29 cases and shows curves for the worst-RMSE case (215 µV).Source: Figure 3 · p. 6
- In a preliminary remapping study, six participants had two to six contacts deactivated in regions predicted to have the lowest neural health.Participants began with at least one year of stable clinical-map use, but follow-up measurements were not collected at consistent timepoints.Source: C. Experiments · p. 7
III. RESULTS
The digital twins represent neural health and stimulation thresholds across fiber bundles, predict clinical SOE responses, and yield neural-health scores correlated with speech recognition. Preliminary remapping results suggest improvement without substantial performance declines in the six reported cases.
Digital twin representation
- The model represents each example case with T(f, S), the stimulation threshold for activating fibers in each bundle, and H_f, an estimate of healthy fibers per bundle.The color-mapped representation makes likely overlap between adjacent electrodes' stimulated fiber-bundle populations visible.Source: §III · p. 7
Experiment 1
- Median RMSE for predicted SOE eCAP measurements was below 50 µV; the worst-case RMSE was 215 µV, with reasonably good agreement in curve shape and magnitude.The worst case is shown in Figure 3.Source: §III · p. 7
Experiment 2
- R=0.78 for CNC word recognition in 29 subjects and R=0.75 for AzBio sentence recognition in noise in 21 subjects, correlated with digital-twin neural health scores.The word scores were measured at six months for 22 subjects, with seven later measurements; sentence scores were measured at six months for 16 subjects, with five later measurements.Source: §III · p. 7
Experiment 3
- 4 of 6 subjects showed uniform CNC word-score improvements after deactivation of 2–6 of 22 electrodes; no substantial decline was observed in any case.All subjects preferred the new map and elected to keep it long term, although follow-up timing was inconsistent.Source: §III · p. 7
IV. CONCLUSIONS
The pipeline builds patient-specific electro-neural-interface digital twins from imaging and electrophysiology, then tests their predictions against clinical neural responses and speech outcomes. Reported findings include neural-health associations with speech recognition and preliminary gains after remapping.
IV. CONCLUSIONS
- The twins model patient-specific spatial, electrical, and neural-health properties using medical images and electrophysiological measurements, with indirect validation against clinical neural responses and speech outcomes.Direct validation would require post-mortem histological analysis.Source: §IV · pp. 7–8
- The digital twins showed good agreement with clinically acquired SOE eCAP measurements, suggesting they capture overlapping stimulation patterns across electrodes.Source: §IV · p. 8
- 50-60% of speech-recognition variability was explained by the neural health score in both quiet and noisy conditions in this dataset.The authors describe this as the strongest single correlating factor reported in the literature, to their knowledge.Source: §IV · p. 8
- Accounting for spatial and tissue-resistivity contributions to eCAPs allows neural health to explain remaining variability in AGF functions and supports objective measurement-based programming.The authors identify possible benefit for pediatric recipients or others who have difficulty providing subjective feedback.Source: §IV · p. 8
- 10 percentage points: CNC scores improved on average in a preliminary prospective remapping study of six subjects, with five ending above baseline and one unchanged.Figure 5 shows CNC word-recognition progression, with a unique color assigned to each subject.Source: Figure 5 · p. 8, §IV · p. 8
- The population model assumes synchronous addition of N1P2 amplitudes, likely underestimates healthy-fiber counts, models only steady-state activation, and was tested on one manufacturer's implants.Source: §IV · p. 8