Lahbib Hana, Mandereau-Bruno Laurence, Goria Sarah, Wagner Vérène, Torres Marion J, Féart Catherine, Helmer Catherine, Pérès Karine, Carcaillon-Bentata Laure
Santé publique France, French National Public Health Agency, Saint-Maurice, 94410, France.
Bordeaux Population Health Research Center, Bordeaux University, INSERM, UMR U1219, Bordeaux, 33000, France.
Sci Rep. 2025 Apr 2;15(1):11344. doi: 10.1038/s41598-025-95629-z.
This study aimed to build a predictive model to identify frailty in the French national health data system (SNDS) so as to create a new tool to monitor and anticipate the disability burden associated with population ageing. We developed the model using the 2012 wave of the French Health, Healthcare, and Insurance Survey (ESPS) linked to the SNDS (n = 2,829). This survey used Fried's frailty phenotype as the gold standard. We compared two statistical approaches - logistic regressions (stepwise and LASSO selection) and random forest - to predict frailty probability based on different SNDS healthcare claims. We indirectly validated the model by comparing (1) the predicted frailty prevalence in the overall French population in the SNDS with the expected prevalence and (2) the predictive ability of the model for 6-year mortality with that of Fried's frailty phenotype. Logistic regression with LASSO selection was retained as the best method to predict frailty. After stratification for age, we obtained three models for individuals aged 55-64, 65-74, and ≥ 75 years (AUC = 0.61, 0.76, and 0.80 respectively). Applying these models to the SNDS, frailty prevalence was comparable to expected prevalence in all sex and age groups: overall prevalence = 12.9% (95%CI: 12.9-12.9) in the SNDS versus 12.0% (95%CI: 10.8-13.2) in the ESPS. Predicted frailty probabilities in the SNDS showed a similar strong association with 6-year mortality (HR=2.6, 95%CI: 1.9-3.5) compared with Fried's phenotype (HR=2.9, 95%CI: 2.1-3.8). Our predictive models are thus useful for estimating frailty probability in the SNDS.
本研究旨在构建一个预测模型，以识别法国国家卫生数据系统（SNDS）中的衰弱情况，从而创建一种新工具，用于监测和预测与人口老龄化相关的残疾负担。我们使用与SNDS相关联的2012年法国健康、医疗保健和保险调查（ESPS）（n = 2829）开发了该模型。这项调查使用弗里德衰弱表型作为金标准。我们比较了两种统计方法——逻辑回归（逐步和LASSO选择）和随机森林——以根据不同的SNDS医疗保健索赔预测衰弱概率。我们通过比较（1）SNDS中法国总体人群的预测衰弱患病率与预期患病率，以及（2）该模型对6年死亡率的预测能力与弗里德衰弱表型的预测能力，间接验证了该模型。保留LASSO选择的逻辑回归作为预测衰弱的最佳方法。按年龄分层后，我们得到了针对55 - 64岁、65 - 74岁和≥75岁个体的三个模型（AUC分别为0.61、0.76和0.80）。将这些模型应用于SNDS，所有性别和年龄组的衰弱患病率与预期患病率相当：SNDS中的总体患病率为12.9%（95%CI：12.9 - 12.9），而ESPS中的为12.0%（95%CI：10.8 - 13.2）。与弗里德表型（HR = 2.9，95%CI：2.1 - 3.8）相比，SNDS中预测的衰弱概率与6年死亡率也显示出类似的强关联（HR = 2.6，95%CI：1.9 - 3.5）。因此，我们的预测模型有助于估计SNDS中的衰弱概率。
