现代隐私约束下构建母-父-婴 linkages 的分层算法
本研究利用台湾全民健康保险研究数据库与出生证明申请数据库(2012-19),开发并验证了一种分层确定性算法,用于从常规行政数据中构建母婴和父子配对。以历史妇幼数据库为参考标准(1,570,007 对母婴、1,510,578 对父子),算法识别母婴配对 1,554,179 对(阳性预测值 97.01%,95%CI 96.99-97.04;灵敏度 96.03%,95%CI 96.00-96.06),父子配对 1,036,980 对(阳性预测值 93.64%,95%CI 93.59-93.69;灵敏度 64.28%,95%CI 64.21-64.36)。母婴性能在 8 年间保持稳定,父子灵敏度存在随时间漂移,需在未来出生队列中重新校准。
要点
- 分层算法仅用常规行政数据即可高准确度构建母婴配对,阳性预测值 97.01%。
- 父子配对灵敏度较低(64.28%),需注意漏匹配风险。
- 母婴性能稳定,父子灵敏度随时间漂移,需定期重新验证。
核心结论
- 分层确定性算法可在隐私法规限制下从常规行政数据构建母婴和父子配对,支持药物警戒研究。
- 母婴配对性能高且稳定,父子配对阳性预测值较高但灵敏度有限且存在时间漂移。
研究设计与样本
- 基于台湾全民健康保险研究数据库和出生证明申请数据库(2012-19)开发并验证分层确定性算法。
- 采用五步分层母婴程序和单步父子匹配,纳入 dependent-guarantor 关系、父亲人口学代理和兄弟姐妹历史。
- 以历史妇幼数据库为确定性参考标准,包含 1,676,503 名登记婴儿中的 1,570,007 对母婴和 1,510,578 对父子。
主要发现
- 算法识别 1,554,179 对母婴(阳性预测值 97.01%,95%CI 96.99-97.04;灵敏度 96.03%,95%CI 96.00-96.06)。
- 算法识别 1,036,980 对父子(阳性预测值 93.64%,95%CI 93.59-93.69;灵敏度 64.28%,95%CI 64.21-64.36)。
- 母婴性能在 8 年期间保持稳定,父子灵敏度存在随时间漂移。
临床怎么用
- 该框架可作为一种可重复、符合隐私法规的模板,用于构建亲子配对以支持父母药物暴露的药物警戒研究。
- 在新环境中应用时应重新验证算法的操作特性。
- 父子灵敏度漂移需在未来出生队列中重新校准。
局限性
- 父子配对灵敏度较低(64.28%),可能影响父子暴露研究的完整性。
- 父子灵敏度存在时间漂移,需定期重新校准。
- 算法仅在台湾数据中验证,外推至其他医疗系统需重新评估。
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
BACKGROUND: Post-marketing safety surveillance of medications used during pregnancy depends on accurate reconstruction of mother-father-infant triads in administrative healthcare data. However, global privacy regulations increasingly restrict the use of direct familial identifiers, and few parent-infant linkage algorithms have been validated against a reference standard. OBJECTIVE: We aimed to develop and validate a hierarchical deterministic algorithm that constructs mother-infant and father-infant linkages from variables widely available in administrative healthcare data internationally. METHODS: Using Taiwan's National Health Insurance Research Database and Birth Certificate Application database (2012-19), we constructed mother-infant and father-infant linkages through a five-step hierarchical mother-infant procedure and a single-step father-infant match, incorporating dependent-guarantor relationships, paternal demographic proxies, and sibling history. We validated the algorithm against the historical Maternal and Child Health Database, which provided a deterministic reference standard of 1,570,007 mother-infant and 1,510,578 father-infant pairs among 1,676,503 registered infants. RESULTS: The algorithm identified 1,554,179 mother-infant pairs (positive predictive value 97.01%, 95% confidence interval [CI] 96.99-97.04; sensitivity 96.03%, 95% CI 96.00-96.06) and 1,036,980 father-infant pairs (positive predictive value 93.64%, 95% CI 93.59-93.69; sensitivity 64.28%, 95% CI 64.21-64.36). Mother-infant performance remained stable across the 8-year period, whereas paternal sensitivity showed a secular drift that we characterized to support recalibration in future birth cohorts. CONCLUSIONS: This hierarchical framework offers a reproducible, privacy regulation-compatible template for constructing parent-infant linkages to support pharmacovigilance research on parental medication exposures. Operating characteristics should be re-validated when the framework is applied in each new setting. Understanding whether medications taken by mothers or fathers before or during pregnancy can affect their children is essential for safe prescribing. Studying these questions requires linking healthcare records of mothers, fathers, and their children but privacy regulations increasingly limit the direct information researchers can use to build these links, and existing methods are rarely tested against a gold-standard benchmark. In this study, we developed and tested a step-by-step method to connect mothers, fathers, and their children in nationwide healthcare records from Taiwan (2012–19), using only routinely collected information such as insurance registration and address details. Our method correctly identified about 97 of every 100 mother–child pairs and 94 of every 100 father–child pairs, checked against Taiwan’s national birth registry as a benchmark. To our knowledge, the father–child method is among the first to be independently validated against such a benchmark. Because the method relies only on information common to health insurance systems worldwide, researchers, regulators, and drug safety agencies can adapt it to their own healthcare data to continue monitoring the safety of medications used by parents.
引用
- Li BM, Cheng MC, Lai EC, et al. Hierarchical Algorithms for Constructing Mother-Father-Infant Linkages Under Modern Privacy Constraints. Drug Saf. 2026. PMID: 42842185. doi:10.1007/s40264-026-01735-w
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。