Scholarlink

Background <p>Many randomized, placebo-controlled trials suggest that the dose–response relationship between dipeptidyl peptidase-4 (DPP-4) inhibitors and change in HbA1c (ΔHbA1c) is non-linear. However, in the real world, patients with type 2 diabetes (T2D) receive varying doses over time depending on their glycemic control. Therefore, to accurately capture real-world HbA1c variability, longitudinal medical records reflecting these dose changes must be analyzed. Crucially, the analysis must also account for the influence of potential confounding factors on HbA1c. Linear mixed-effects models cannot adequately assess this non-linear relationship. The aim of this study was to evaluate the association between daily dose of DPP-4 inhibitors and ΔHbA1c in the real world by developing and interpreting tree-based machine-learning models with random effects.</p> Methods <p>Longitudinal information on T2D patients was extracted from electronic medical records. Four machine-learning models with random effects were constructed to predict change in HbA1c: linear mixed-effects models (LMMs) with/without a backward elimination method, a mixed-effects random forest model (MERF), and a combining tree-boosting with Gaussian process and mixed-effects model (GPBoost). The MERF and GPBoost were interpreted by SHapley Additive exPlanations and partial dependence.</p> Results <p>Non-linear machine-learning models such as MERF and GPBoost had better predictive performance than LMMs. When interpreting MERF and GPBoost, HbA1c level decreased in a dose-dependent manner within the range of sitagliptin 0–50&#xa0;mg/day, but no difference was observed between 50 and 100&#xa0;mg/day. For teneligliptin, while 10&#xa0;mg/day was suggested to have a slightly weaker HbA1c-lowering effect than 20 or 40&#xa0;mg/day, there was no difference between 20 and 40&#xa0;mg/day. Regarding vildagliptin, 100&#xa0;mg/day showed a tendency to lower HbA1c level more than 50&#xa0;mg/day.</p> Conclusions <p>Non-linear machine-learning models can clearly demonstrate the dose–response relationship of the HbA1c-lowering effects of DPP-4 inhibitors in real-world T2D patients. For sitagliptin and teneligliptin, the HbA1c-lowering effect plateaued within the approved dose range. With regard to vildagliptin, it is unclear whether dose–response saturation was observed due to the limited number of dosing regimens. Our findings support the results of previous trials, and machine-learning models demonstrated that it is possible to evaluate the dose–response relationship in the real world.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Association between daily dose of dipeptidyl peptidase-4 inhibitors and change in glycated hemoglobin in patients with type 2 diabetes: interpretation of mixed-effects machine-learning models using electronic medical records

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月22日

查看全文
  • Takashi Hayakawa,
  • Hayato Akimoto,
  • Takuya Nagashima,
  • Kimino Minagawa,
  • Yasuo Takahashi,
  • Satoshi Asai

摘要

Background

Many randomized, placebo-controlled trials suggest that the dose–response relationship between dipeptidyl peptidase-4 (DPP-4) inhibitors and change in HbA1c (ΔHbA1c) is non-linear. However, in the real world, patients with type 2 diabetes (T2D) receive varying doses over time depending on their glycemic control. Therefore, to accurately capture real-world HbA1c variability, longitudinal medical records reflecting these dose changes must be analyzed. Crucially, the analysis must also account for the influence of potential confounding factors on HbA1c. Linear mixed-effects models cannot adequately assess this non-linear relationship. The aim of this study was to evaluate the association between daily dose of DPP-4 inhibitors and ΔHbA1c in the real world by developing and interpreting tree-based machine-learning models with random effects.

Methods

Longitudinal information on T2D patients was extracted from electronic medical records. Four machine-learning models with random effects were constructed to predict change in HbA1c: linear mixed-effects models (LMMs) with/without a backward elimination method, a mixed-effects random forest model (MERF), and a combining tree-boosting with Gaussian process and mixed-effects model (GPBoost). The MERF and GPBoost were interpreted by SHapley Additive exPlanations and partial dependence.

Results

Non-linear machine-learning models such as MERF and GPBoost had better predictive performance than LMMs. When interpreting MERF and GPBoost, HbA1c level decreased in a dose-dependent manner within the range of sitagliptin 0–50 mg/day, but no difference was observed between 50 and 100 mg/day. For teneligliptin, while 10 mg/day was suggested to have a slightly weaker HbA1c-lowering effect than 20 or 40 mg/day, there was no difference between 20 and 40 mg/day. Regarding vildagliptin, 100 mg/day showed a tendency to lower HbA1c level more than 50 mg/day.

Conclusions

Non-linear machine-learning models can clearly demonstrate the dose–response relationship of the HbA1c-lowering effects of DPP-4 inhibitors in real-world T2D patients. For sitagliptin and teneligliptin, the HbA1c-lowering effect plateaued within the approved dose range. With regard to vildagliptin, it is unclear whether dose–response saturation was observed due to the limited number of dosing regimens. Our findings support the results of previous trials, and machine-learning models demonstrated that it is possible to evaluate the dose–response relationship in the real world.