光华讲坛——社会名流与企业家论坛第6681期

主题：Estimating Patient Health Transition from Data Censored by Treatment-Effect-Based Policies从基于治疗效果的政策所删减的数据中估算患者的健康状况转变

主讲人：郑智超

主持人：肖辉

时间：12月12日 10:00-12:00

举办地点：西南财经大学柳林校区通博楼D301

主办单位：西南财经大学管科学院 科研处

主讲人简介：

Zhichao Zheng is an Associate Professor of Operations Management at the Singapore Management University. His main research interests lie in data analytics and optimization for healthcare operations management and medical decision-making. He also applies his research in sharing economics, supply chain risk management, etc. His research has appeared in Operations Research, Management Science, and Manufacturing & Service Operations Management, among others. He received his BS (First Class Honors) in Applied Mathematics from the National University of Singapore in 2009 and Ph.D. in Management from the Department of Decision Sciences (renamed to Department of Analytics & Operations) at the National University of Singapore in 2013.郑智超是新加坡管理大学运营管理教授。他的主要研究兴趣是医疗运营管理和医疗决策的数据分析与优化。他还将研究成果应用于共享经济学、供应链风险管理等领域。他的研究成果发表在《运营研究》、《管理科学》、《制造与服务运营管理》等杂志上。他于2009年获得新加坡国立大学应用数学学士学位（一等荣誉学位），并于2013年获得新加坡国立大学决策科学系（已更名为分析与运营系）管理学博士学位。

内容简介：

Treatment-effect-based decision policies are increasingly used in healthcare problems. Such policies leverage predictive information on patient health transitions and treatment outcomes for treatment recommendations. However, these policies can significantly censor the observation of patients’ health transitions and distort the estimation of transition probability matrices (TPMs). We propose a structural model to recover the underlying true TPMs from censored transition observations. We show that the estimated TPM from the structural model is consistent and asymptotically normally distributed and also maximizes the log-likelihood of observing the data. Using hypothesized data with known ground truth TPMs, we demonstrate the advantages of our model against benchmark estimation methods that ignore the censoring mechanism. We further implement our model to estimate patient health transitions using observed data for the extubation problems in an intensive care unit (ICU). Formulating the extubation problem as a classical optimal stopping Markov Decision Process model, we show that the proposed method, with more accurate estimated TPMs considering treatment-effect-based policy censoring, can reduce patients’ length of stay in the ICU compared to benchmark methods.基于治疗效果的决策政策越来越多地应用于医疗保健问题。此类政策利用患者健康状况转变和治疗结果的预测信息来提出治疗建议。然而，这些政策会严重删减对患者健康转变的观察，并扭曲对转变概率矩阵（TPM）的估计。我们提出了一个结构模型，以从剔除的过渡观测中恢复潜在的真实TPM。我们的研究表明，从结构模型中估算出的TPM是一致的、渐近正态分布的，而且能使观测数据的对数似然最大化。通过使用具有已知基本事实TPM的假设数据，我们证明了我们的模型与忽略删减机制的基准估计方法相比所具有的优势。我们进一步使用我们的模型，利用观察到的重症监护室（ICU）拔管问题数据来估计病人的健康状况转变。我们将拔管问题表述为一个经典的最优停止马尔可夫决策过程模型，结果表明，与基准方法相比，考虑到基于治疗效果的政策剔除，所提出的方法具有更精确的估计TPM，可以缩短患者在重症监护室的住院时间。