文献速递耳科与耳神经小儿耳鼻咽喉

从二分类到个性化:一种用于人工耳蜗候选资格的新型机器学习概率评分

Binary to Personalized: A Novel Machine Learning Probability Score for Cochlear Implant Candidacy.
PubMed 文献 · Laryngoscope · 2026-08-21

本研究基于机构人工耳蜗登记库和HERMES数据库(1878例患者,3756耳),开发并验证机器学习模型预测人工耳蜗候选资格,输出个性化概率评分。模型以耳特异性助听后CNC词得分<50%和双耳最佳助听AzBio句得分<60%为候选标准,与60/60规则(PTA≥60 dB;WRS≤60%)进行基准比较。结果显示机器学习模型在敏感性及部分F1、特异性上具有互补价值,尤适用于临界「灰色地带」患者的个性化咨询与共享决策。

要点

  • 机器学习模型预测CNC<50%时敏感性0.80-0.84,高于60/60规则的0.77。
  • 预测AzBio<60%时机器学习F1为0.79-0.80,优于60/60规则的0.71。
  • 模型输出校准良好的个体化概率,有助于临界患者的个性化咨询。

核心结论

  • 机器学习模型虽未显著优于60/60规则,但可识别可能被遗漏的候选者并提供个体化概率评分,支持共享决策。

研究设计与样本

  • 基于机构CI登记库的回顾性队列研究及HERMES数据库,共1878例患者、3756耳。
  • 采用逻辑回归、随机森林和XGBoost三种监督机器学习分类模型,按TRIPOD-AI指南与60/60规则进行基准比较。
  • 数据按80/20拆分,采用嵌套10折交叉验证调参,并应用SMOTE评估类别不平衡影响。

主要发现

  • 对于CNC<50%标准,60/60规则F1分数更高(0.83 vs. 0.77-0.79),但机器学习模型敏感性更高(0.80-0.84 vs. 0.77),AUROC达0.88-0.89。
  • 对于AzBio<60%标准,机器学习模型在F1分数(0.79-0.80 vs. 0.71)、精确率(0.77-0.78 vs. 0.68)和特异性(0.86-0.87 vs. 0.76)上均优于60/60规则,AUROC为0.89-0.90。
  • 校准分析证实模型输出校准良好、具有临床可解释性的概率分布。
  • SMOTE平衡数据集未改变模型性能。

关键数值与推荐

  • 人工耳蜗候选标准1:耳特异性助听后CNC词得分<50%。
  • 人工耳蜗候选标准2:双耳最佳助听AzBio句得分<60%。
  • 对照规则:60/60规则(PTA≥60 dB;WRS≤60%)。

临床怎么用

  • 机器学习概率评分可作为辅助工具,帮助识别传统规则可能遗漏的候选者,并为临界患者提供个体化咨询依据。
  • 模型输出的概率分布经过校准,可用于支持医患共享决策。

局限性

  • 机器学习模型并未显著优于60/60规则,其临床增量价值需进一步验证。
  • 本研究为回顾性队列研究,外部泛化性有待前瞻性多中心验证。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

OBJECTIVE: To develop and validate a machine learning (ML)-based screening tool for cochlear implant (CI) candidacy that generates personalized probability scores and supports individualized, patient-centered counseling. METHODS: Institutional CI registry retrospective cohort study and HERMES database (1878 patients, 3756 ears). Supervised ML classification models predicted CI candidacy: (1) ear-specific aided CNC word score < 50%, and (2) bilateral best-aided AzBio sentence scores < 60%. Three ML models (logistic regression, random forest, XGBoost) were benchmarked against the 60/60 rule (PTA ≥ 60 dB; WRS ≤ 60%) per TRIPOD-AI guidelines. Input features included audiometric thresholds, word recognition scores, age. Data were split 80/20, hyperparameter tuning via nested 10-fold cross-validation. Performance metrics included sensitivity, specificity, AUROC, F1 score, and calibration curves. SMOTE was applied to evaluate class imbalance effects. A web-based platform was deployed for user-friendly access. RESULTS: For CNC < 50%, the 60/60 rule achieved a higher F1 score (0.83 vs. 0.77-0.79), but ML models demonstrated higher sensitivity (0.80-0.84 vs. 0.77) and excellent AUROC (0.88-0.89). For AzBio < 60%, ML models outperformed the 60/60 rule in F1 score (0.79-0.80 vs. 0.71), precision (0.77-0.78 vs. 0.68), and specificity (0.86-0.87 vs. 0.76), with AUROC 0.89-0.90. Calibration analysis confirmed well-calibrated, clinically interpretable probability distributions. SMOTE-balanced datasets did not alter performance. CONCLUSION: Although ML models are not dramatically superior to the 60/60 rule, they offer complementary value by identifying candidates who might otherwise be missed and providing individualized probability scores for both CNC and AzBio CI candidacy criteria. These outputs enable more personalized counseling and shared decision-making, particularly for patients in the borderline "gray zone".

来源:Laryngoscope,查看原文。版权归原作者所有。

引用

  1. Shew MA, Jawad M, Pavelchek C, et al. Binary to Personalized: A Novel Machine Learning Probability Score for Cochlear Implant Candidacy. Laryngoscope. 2026. PMID: 42629619. doi:10.1002/lary.70845

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。