数据驱动策略用于卡比马唑剂量调整:探索机器学习在甲状腺功能亢进控制中的解决方案
本研究基于UHD 353例自身免疫性甲状腺功能亢进患者数据,测试多种机器学习分类算法以预测卡比马唑剂量调整。随机森林表现最佳,平均F1分数0.731;所选模型F1=0.751、Brier分数=0.38。模拟使用场景显示,若投入实践预期可提高准确性。
要点
- 随机森林方法表现最佳,平均F1分数为0.731。
- 所选模型基于预测准确性(F1=0.751)与模型置信度(Brier分数=0.38)的平衡评估。
- 模拟使用场景显示,若模型投入实践,预期可提高准确性。
核心结论
- 模拟使用场景评估预测误差随时间累积,确定若模型投入实践预期可提高准确性。
研究设计与样本
- 从UHD 421例甲状腺功能亢进患者中提取并匿名化数据,353例(83.85%)纳入研究。
- 测试不同机器学习分类算法在多种数据处理方案下的表现。
- 采用迭代方法,先进行初始模型选择,再进行特征选择以提升性能。
- 使用加权F1分数和Brier分数评估模型,以选择置信度最高的最佳模型。
主要发现
- 随机森林方法取得最佳性能,平均F1分数为0.731。
- 所选模型基于预测准确性(F1=0.751)与模型置信度(Brier分数=0.38)的平衡评估。
临床怎么用
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
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: University Hospitals Dorset (UHD) has more than 1000 thyroid patient contacts annually. These are primarily patients with autoimmune hyperthyroidism treated with carbimazole titration. Dose adjustments are made by a healthcare professional (HCP) based on the results of thyroid function tests, who then prescribes a dose and communicates this to the patient via letter. This is time consuming and introduces treatment delays. This study aimed to replace some time-intensive manual dose adjustments with a machine learning model to determine carbimazole dosing. This can in the future serve patients with rapid and safe dose determination and ease the pressures on HCPs. METHODS: Data from 421 hyperthyroidism patients at UHD were extracted and anonymized. A total of 353 patients (83.85%) were included in the study. Different machine learning classification algorithms were tested under several data processing regimes. Using an iterative approach, consisting of an initial model selection followed by a feature selection method, the performance was improved. Models were evaluated using weighted F1 scores and Brier scores to select the best model with the highest confidence. RESULTS: The best performance is achieved using a random forest (RF) approach, resulting in good average F1 scores of 0.731. A model was selected based on a balanced assessment considering the accuracy of the prediction (F1 = 0.751) and the confidence of the model (Brier score = 0.38). CONCLUSION: To simulate a use-case, the accumulation of the prediction error over time was assessed. It was determined that an improvement in accuracy is expected if this model was to be deployed in practice.
引用
- Reich T, Bakirov R, Budka D, et al. Data-Driven Strategies for Carbimazole Titration: Exploring Machine Learning Solutions in Hyperthyroidism Control. J Clin Endocrinol Metab. 2025. PMID: 39271154. doi:10.1210/clinem/dgae642
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。