一种基于神经反应遥测和电诱发听性脑干反应阈值的双数据驱动方法用于预测人工耳蜗动态范围参数
A dual-data-driven approach for predicting cochlear implant dynamic range parameters using neural response telemetry and electrically evoked auditory brainstem response thresholds.
文献信息
| PMID | 42874450 |
|---|---|
| 原文 | 在 PubMed 查看原文 ↗ |
| 发表日期 | 2026 |
| 作者 | Yao Liu |
| 作者单位 | Department of Otorhinolaryngology Head and Neck Surgery, Tianjin First Central Hospital, China; Institute of Otolaryngology of Tianjin, China; Key Laboratory of Auditory Speech and Balance Medicine, China; Key Medical Cultivation Discipline of Tianjin (Otolaryngology), China; Quality Control Centre of Otolaryngology, Tianjin, 300192, China. |
| 期刊 | Hearing research |
| SCI 分区 | Q1 |
| IF | 3.2 |
| 研究类型 | AI/ML · 临床 |
| 所属专科 | 耳科 |
中文摘要
背景： 目前，结合神经反应遥测（NRT）和电诱发听性脑干反应（EABR）来预测人工耳蜗（CI）拟合参数的临床研究仍然很少。
目的： 探讨CI接受者中NRT阈值、EABR阈值与主观行为阈值T/C级之间的相关性，并开发一种基于人工神经网络的CI拟合参数预测模型。
方法： 从37名CI接受者中收集电极1、6、11、16和22的NRT阈值、EABR阈值以及主观行为阈值T/C级，随后进行相关性分析。以NRT和EABR阈值作为联合输入特征，构建了一种结合贝叶斯正则化（BR）机制的反向传播（BP）神经网络（BR-BP），用于预测T级和C级。
结果： NRT和EABR阈值均与T级和C级呈显著正相关（P < 0.05）。直接结合NRT和EABR阈值的三维线性预测模型因严重的多重共线性而失败，最大平均绝对误差（MAE）高达373.86 CL。相比之下，BR-BP神经网络成功实现了非线性特征解耦，有效缓解了小样本常见的过拟合问题，并将所有测试电极的预测误差稳定限制在约60 CL以内。
结论： NRT和EABR阈值与主观行为CI拟合参数存在显著相关性。虽然单源预测模型有时可能在某些电极上提供局部准确性，但其易受信号丢失影响的脆弱性限制了其临床安全性。所提出的神经网络模型由双客观生理指标驱动，有效克服了传统线性预测和单源依赖的局限性。通过优先考虑全局容错而非边缘局部精度，该方法为CI编程提供了一种高度稳健、安全且可靠的自动化起始映射，对不合作患者尤其有价值。
英文摘要
BACKGROUND: Currently, clinical research combining Neural Response Telemetry (NRT) and Electrically Evoked Auditory Brainstem Responses (EABR), for the prediction of cochlear implant (CI) fitting parameters remains scarce.
OBJECTIVES: To investigate the correlations between NRT thresholds, EABR thresholds, and subjective behavioral threshold T/C-levels in CI recipients, and to develop an artificial neural network-based prediction model for CI fitting parameters.
METHODS: NRT thresholds, EABR thresholds, and subjective behavioral threshold T/C-levels across electrodes 1, 6, 11, 16, and 22 were collected from 37 CI recipients, followed by correlation analyses. Using NRT and EABR thresholds as combined input features, a backpropagation (BP) neural network integrated with a Bayesian regularization (BR) mechanism (BR-BP) was constructed to predict the T- and C-levels.
RESULTS: Both NRT and EABR thresholds exhibited significant positive correlations with T- and C-levels (P < 0.05). A three-dimensional linear prediction model directly combining NRT and EABR thresholds failed due to severe multicollinearity, yielding a maximum mean absolute error (MAE) of up to 373.86 CL. In contrast, the BR-BP neural network successfully achieved nonlinear feature decoupling, effectively mitigating the overfitting commonly associated with small sample sizes, and stably constrained the prediction errors across all tested electrodes to within approximately 60 CL.
CONCLUSIONS: NRT and EABR thresholds exhibit significant correlations with subjective behavioral CI fitting parameters. While single-source predictive models may occasionally offer localized accuracy on certain electrodes, their vulnerability to signal loss limits their clinical safety. The proposed neural network model, driven by dual objective physiological metrics, effectively overcomes the limitations of traditional linear prediction and single-source dependency. By prioritizing global fault tolerance over marginal local precision, this approach provides a highly robust, safe, and reliable automated starting map for CI programming, proving especially valuable for uncooperative patients.