A machine learning model can detect when a cochlear implant electrode array has slipped out of position using only the impedance measurements that are already collected during routine follow-up, according to a retrospective study published in the Journal of Medical Systems. The approach correctly identified electrode array migration in 14 of 16 cases with radiologically confirmed displacement, while flagging none of the 16 stable control cases, yielding a sensitivity of 88 percent and a specificity of 100 percent. Because the method relies on telemetry data streamed directly from the implant, it requires no additional imaging and no ionizing radiation, a property the researchers argue could make continuous screening practical for the first time.
Electrode array migration is a recognized but uncommon complication of cochlear implantation, with reported incidence estimates ranging from 3 to 7 percent. When the array moves, the effective frequency allocation of the implant shifts, and displacements as small as one millimeter can produce perceptible changes for patients. In the worst cases, migration necessitates revision surgery, which the authors note is estimated to account for roughly 1 to 15 percent of all revision procedures. Yet migration often goes undetected during standard follow-up, particularly when symptoms are mild, because comparative radiological imaging is not performed systematically. Current detection is largely reactive: clinicians typically suspect a problem only after nonspecific signs such as loss of neural response telemetry, declining speech understanding, device rejection, or unpleasant stimulation effects.
The research team, drawn from four cochlear implant centers, assembled a cohort of 32 cases: 16 with radiologically confirmed electrode array migration and 16 matched stable controls, matched on electrode array type, age at surgery, sex, and implanted ear side. The cases corresponded to 28 patients, including four bilateral implantations, with 13 females and 15 males. Patient ages ranged from 1 to 89 years, with a median of 56, and there were 15 left-sided and 17 right-sided implantations. All patients received lateral wall electrode arrays from a single manufacturer, implanted between 2011 and 2025, and each case required at least two radiological scans of the same ear at different postoperative time points. Cases with abnormal cochlear conditions such as otosclerosis, enlarged vestibular aqueduct, or Ménière’s disease were excluded.
Ground truth was established by comparing the first and second postoperative scans, using three-dimensional CT or cone beam CT in 20 cases and X-ray projections in 12. Migration was defined as an absolute longitudinal displacement of the most basal electrode exceeding 1.0 millimeter, a threshold chosen to exceed the expected measurement noise of the radiological reference, given that training-set CT scans had an average in-plane resolution of 0.2 millimeters and average slice spacing of 0.27 millimeters. The threshold is roughly half an inter-electrode spacing for the most common array in the dataset, which has an active stimulation range of 23.1 millimeters across 12 electrodes. Transverse displacement toward the modiolar or lateral wall was not assessed.
The estimation model itself was trained on a separate cohort of 179 cases from a single center, all requiring preoperative and postoperative CT scans. Features were extracted from the full transimpedance matrix recorded during surgery, cochlear dimensions measured on preoperative CT, and implant parameters including stimulation current and ground path impedance. The target was the linear insertion depth of the most basal electrode, labeled manually from co-registered CT scans along the modiolar axis. Nested cross-validation with ten outer folds and five inner folds compared six regression approaches; a random forest regressor using 15 features won, predicting basal electrode position with a mean absolute error of 0.94 millimeters plus or minus 0.75, with errors ranging from −3.5 to 2.4 millimeters. The worst performer, support vector regression with 30 features, reached only 1.0 plus or minus 0.78 millimeters, a difference the authors interpret as evidence that the information content of the impedance features, not algorithm choice, set the performance ceiling. To prevent data leakage, separate models were retrained for each of the 32 study cases with that case excluded from training.
For detection, each impedance recording produced one position estimate, which was compared against a running baseline starting with the first, usually intraoperative, recording. A signed change exceeding the 1-millimeter threshold flagged migration, with negative changes classified as emigration out of the cochlea and positive changes as immigration deeper in. On average, the first migration event was detected 10.6 months after surgery, ranging from 0.2 to 41.7 months. The model caught migration in clinically inconspicuous cases with largely unremarkable impedance trajectories, and in long-term follow-up spanning more than 12 years. In one illustrative emigration case, the model flagged the event 13 months after surgery, 13 months before radiological confirmation at month 26, and well before visual inspection of impedances would have raised suspicion, when the most basal electrode exceeded 20 kilo-ohms at month 37.
The study also reports what the authors describe as the first documentation of electrode immigration, in which the array advances further into the cochlea after surgery. Radiologically confirmed immigration occurred in 2 of the 16 migration cases, with the most basal electrode displaced by +1.4 and +1.6 millimeters. The authors suggest immigration may occur when the lead cable is coiled too tightly within the mastoid cavity, so that release of stored bending energy pushes the array deeper. Although immigration may have less severe audiological consequences because the electrodes remain inside the cochlea, it can still shift stimulation sites relative to the intended frequency allocation and potentially impair speech understanding.
Group-level impedance analysis revealed a systematic spatial pattern at the first model-detected migration event. Impedances were comparable between groups at the intraoperative recording, but at migration detection, control cases showed slightly higher impedances at apical electrodes while migration cases showed slightly higher impedances at the most basal electrodes, along with greater variability there. Only two of the comparisons across all 12 electrodes and both time points reached statistical significance. The strongest was at the medial electrode E5, with median impedances of 4.1 versus 6.9 kilo-ohms in migration and control cases, a median difference of 2.8 kilo-ohms with a 95 percent confidence interval of 0.3 to 3.3, and a Mann-Whitney p-value of 0.014. The authors caution that these patterns were not consistently present in every migration case, which is precisely why simple visual inspection of impedance curves is insufficient.
The method has clear limits. Two migration cases were missed; in both, displacement of the most basal electrode was less than one inter-electrode spacing, at −1.6 and −1.3 millimeters for arrays with 2.1-millimeter spacing, apparently producing too little impedance change to detect. Direction classification was imperfect, with some confusions attributed to a training set dominated by emigration cases. Temporal accuracy was not assessed, since impedance data were recorded at regular clinical intervals while CT scans were obtained only on clinical indication. The training cohort, though expanded to 179 cases, came from a single center and was dominated by one array type, with 162 of the cases using the same FLEX28 array, so generalizability across array types and centers remains unproven. Impedance changes from inflammatory and foreign-body reactions, unrelated to electrode position, also constitute a confounding signal, although most model features relied on off-diagonal transimpedance values that are less affected by such effects.
The authors position the tool as a clinical decision support system intended to notify clinical engineers and audiologists that further review may be warranted, not as an autonomous diagnostic. Because manufacturers already provide remote impedance measurement, integration into telehealth workflows appears technically straightforward, and the approach could enable large-scale retrospective screening of existing implant databases to produce a less biased estimate of migration incidence. Presented as a first proof-of-concept, the method requires validation on larger, external cohorts, including comparisons with clinician performance, before it can be integrated into standard post-implantation care. The researchers also suggest that secure fixation of the implant’s internal lead during surgery may reduce migration risk from a preventive standpoint, and that future work could pair impedance telemetry with electrically evoked compound action potentials to improve robustness.
Subject of Research: Machine learning detection of cochlear implant electrode array migration using routine impedance telemetry
Article Title: Automated, Radiation-Free Detection of Cochlear Implant Electrode Array Migration Using Routine Impedance Telemetry
Article References: Schraivogel, S., Rader, T., Weissgerber, T., Caversaccio, M., & Wimmer, W. (2026). Automated, Radiation-Free Detection of Cochlear Implant Electrode Array Migration Using Routine Impedance Telemetry. Journal of Medical Systems, 50(1), Article 158. https://doi.org/10.1007/s10916-026-02468-0
Image Credits: AI Generated
DOI: 10.1007/s10916-026-02468-0
Keywords: cochlear implant, electrode array migration, impedance telemetry, machine learning, random forest, radiation-free screening, pediatric patients, revision surgery, electrode immigration, telehealth, retrospective study, Journal of Medical Systems
News Source: Ophelia Keating. (November 5, 2026). Implant Telemetry Spots Shifting Cochlear Electrodes Without X-Rays. Scienmag.
