Page 111 - 《中国药房》2025年19期
P. 111

3.3 本研究的局限性                                             taset[J]. Sci Data,2023,10(1):1.
              本研究在多分类任务评估中存在一定局限性:(1)                        [ 8 ]  MRAMBA L K,LIU X,LYNCH K F,et al. Detecting po‐
          研究采用了 ADASYN 过采样技术,其效果依赖于少数                             tential  outliers  in  longitudinal  data  with  time-dependent
          类的原始分布,若少数类样本过少或特征重叠严重,可                                covariates[J]. Eur J Clin Nutr,2024,78(4):344-350.
                                                             [ 9 ]  YAN F J,CHEN X H,QUAN X Q,et al. Association be‐
          能生成与多数类重叠的“虚假样本”,导致分类器混淆,
                               [15]
          降低少数类的分类精度 。当类别之间存在重叠时,                                 tween the stress hyperglycemia ratio and 28-day all-cause
                                                                  mortality in critically ill patients with sepsis:a retrospec‐
          OvR-AUC 评估方法的独立性假设可能不成立,从而导
                                                                  tive  cohort  study  and  predictive  model  establishment
          致 AUC 虚高,无法以 AUC 反映类别特定的误分类代价                           based on machine learning[J]. Cardiovasc Diabetol,2024,
              [16]
          差异 。(2)特征工程未能充分表征浓度梯度变化规律,                              23(1):163.
          导致低、中浓度区间区分能力受限。针对这一问题,可                           [10]  PRIYADHARSHINI M,BANU A F,SHARMA B,et al.
          通过构建时序相关特征(如剂量-效应矩阵),引入药物                               Hybrid multi-label classification model for medical appli‐
          动力学相关的时序特征来提升模型对动态浓度变化的                                 cations  based  on  adaptive  synthetic  data  and  ensemble
          捕捉能力 。(3)一些潜在的混杂因素(如遗传多态性、                              learning[J]. Sensors,2023,23(15):6836.
                  [17]
          合并用药、合并症或其他疾病表型等)可能对预测结果                           [11]  CHICCO  D,JURMAN  G.  The  advantages  of  the  Mat‐
                  [18]
          产生影响 ,但这些因素均未被本研究纳入,可能造成                                thews  correlation  coefficient (MCC)  over  F1  score  and
          系统偏差。(4)单一模型可能难以全面捕捉数据中的复                               accuracy in binary classification evaluation[J]. BMC Ge‐
          杂模式,尤其在面对高维、非线性及多模态数据时表现                                nomics,2020,21(1):6.
          受限,深度学习模型、图神经网络等先进架构有望弥补                           [12]  SWANSON K,WU E,ZHANG A,et al. From patterns to
                                                                  patients:advances in clinical machine learning for cancer
                  [19]
          这一短板 。(5)模型的验证目前仅基于内部数据和交                               diagnosis,prognosis,and  treatment[J].  Cell,2023,186
          叉验证,尚未在外部独立数据集中进行测试,其泛化能                               (8):1772-1791.
          力仍有待进一步验证。(6)尽管本研究中的 RF 和 Tab‐                     [13]  TRAN V Q,BYEON H. Predicting dementia in Parkinson’s
          PFN模型在分类性能上表现出色,但其“黑箱”特性限制                              disease on a small tabular dataset using hybrid LightGBM-
          了对预测机制的深入理解。                                            TabPFN and SHAP[J]. Digit Health,2024,10:1-15.
          4 结语                                               [14]  HUI K H M,LUI C Y G,WU K L A,et al. Multi-center
              机器学习模型 RF 和 TabPFN 在万古霉素血药谷浓                        prospective population pharmacokinetic study and the per‐
          度分类预测中表现出一定优势,但在低、中浓度类别上                                formance of web-based individual dose optimization appli‐
          的表现仍有改进空间。未来研究应优化试验设计与特                                 cation  of  intravenous  vancomycin  for  adults  in  Hong
                                                                  Kong:a  study  protocol[J].  PLoS  One,2022,17(5):
          征工程,拓展多中心数据,验证模型的适用性,提升模型
                                                                  e0267894.
          性能和可解释性,推动机器学习在TDM中的实际应用。                          [15]  PEREIRA T,ABBASI M,OLIVEIRA J L,et al. Optimi-
          参考文献                                                    zing  blood-brain  barrier  permeation  through  deep  rein‐
          [ 1 ]  WILLIAMS P G,TABAH A,COTTA M O,et al. Interna‐   forcement learning for de novo drug design[J]. Bioinfor‐
              tional survey of antibiotic dosing and monitoring in adult   matics,2021,37(Suppl. 1):i84-i92.
              intensive care units[J]. Crit Care,2023,27(1):241.  [16]  MBIZVO  G  K,LARNER  A  J.  Receiver  operating
          [ 2 ]  LIM A S,FOO S H W,BENJAMIN SENG J J,et al. Area-  characteristic  plot  and  area  under  the  curve  with  binary
              under-curve-guided  versus  trough-guided  monitoring  of   classifiers:pragmatic  analysis  of  cognitive  screening
              vancomycin and its impact on nephrotoxicity:a systematic   instruments[J].  Neurodegener  Dis  Manag,2021,11(5):
              review  and  meta-analysis[J]. Ther  Drug  Monit,2023,45  353-360.
              (4):519-532.                                   [17]  HUANG X H,YU Z,WEI X,et al. Prediction of vanco‐
          [ 3 ]  ANGELOPOULOS  A  N,BATES  S,FANNJIANG  C,        mycin dose on high-dimensional data using machine lear-
              et al. Prediction-powered inference[J]. Science,2023,382  ning techniques[J]. Expert Rev Clin Pharmacol,2021,14
              (6671):669-674.                                    (6):761-771.
          [ 4 ]  HUANG X H,YU Z,BU S H,et al. An ensemble model   [18]  YIN  M  H,JIANG Y  L,YUAN Y  W,et  al.  Optimizing
              for prediction of vancomycin trough concentrations in pedia-   vancomycin  dosing  in  pediatrics:a  machine  learning
              tric patients[J]. Drug Des Devel Ther,2021,15:1549-1559.  approach  to  predict  trough  concentrations  in  children
          [ 5 ]  OTA R,YAMASHITA F. Application of machine learning   under four years of age[J]. Int J Clin Pharm,2024,46(5):
              techniques to the analysis and prediction of drug pharma‐  1134-1142.
              cokinetics[J]. J Control Release,2022,352:961-969.  [19]  ABBAS  Q,DAADAA  Y,RASHID  U,et  al.  HDR-
          [ 6 ]  HOLLMANN N,MÜLLER S,PURUCKER L,et al. Ac‐        EfficientNet:a classification of hypertensive and diabetic
              curate predictions on small data with a tabular foundation   retinopathy  using  optimize  EfficientNet  architecture[J].
              model[J]. Nature,2025,637(8045):319-326.            Diagnostics,2023,13(20):3236.
          [ 7 ]  JOHNSON  A  E  W,BULGARELLI  L,SHEN  L,et  al.             (收稿日期:2025-04-21  修回日期:2025-08-01)
              MIMIC-Ⅳ,a freely accessible electronic health record da‐                            (编辑:刘明伟)


          中国药房  2025年第36卷第19期                                              China Pharmacy  2025 Vol. 36  No. 19    · 2453 ·
   106   107   108   109   110   111   112   113   114   115   116