Deep learning for representation of codes used for SEER-Medicare claims research

 描述(由申请人提供)：我们建议开发一个算法和用户友好的软件，以更好地识别使用联邦医疗保险索赔数据的治疗。我们将使用监测、流行病学和最终结果(SEER)数据库中列出的程序作为黄金标准来验证我们的方法。通过这种方式，我们希望更好地将使用联邦医疗保险索赔数据确定的程序与SEER列出的程序相匹配。本研究的重点是观察性(即非随机化)数据。运行良好的随机临床试验可以提供最佳水平的治疗效果证据。然而，美国的随机试验在许多干预措施中收效甚微。尽管设计良好的随机临床试验应该是金标准，但设计良好的观察性研究可能是获得关于某些癌症干预措施的比较有效性的推断的唯一方法。在癌症研究中，最常用的观察性研究数据库之一是链接的SEER-Medicare数据库。SEER-Medicare的数据为许多癌症疗法的有效性提供了有用的衡量标准。使用医疗保险数据识别相关治疗和诊断代码的算法通常基于临床推理和科学证据。例如，一组研究人员开发了一种算法，用于在肾癌病例中识别腹腔镜手术，然后才很好地开发了腹腔镜手术的索赔代码。虽然这样的算法对其他进行类似研究的人很有用，但在SEER癌症登记确定的治疗和通过联邦医疗保险索赔确定的治疗之间可能仍然存在严重的不匹配。在这项工作中，我们建议开发一种严格的机器学习算法，可以帮助研究人员更好地识别医疗保险索赔数据中的治疗方法。具体地说，我们将设计一个神经语言建模算法，并实现一个找到诊断和过程代码的矢量表示的软件系统。我们计划使用神经语言建模算法从SEER-Medicare索赔数据中学习向量表示，在这些数据中，相关程序和诊断代码是“相邻的”(即密切相关)。我们将调查我们在社区内确定的代码是否与发表的SEER-Medicare研究所使用的程序代码相对应。然后，我们将设计一个软件助手界面，允许研究人员探索哪些代码与给定的诊断种子或程序代码相关。最后，我们将通过将使用Medicare Claims识别的程序与SEER数据库中列出的程序进行比较，来调查该算法的敏感性和特异性。我们将复制一篇发表的SEER-Medicare论文中的分析，以调查使用我们的新算法与使用发表的论文中的算法时，估计的治疗效果是否不同。

 DESCRIPTION (provided by applicant): We propose developing an algorithm and user-friendly software to better identify treatments using Medicare claims data. We will validate our approach using procedures listed in the Surveillance, Epidemiology, and End Results (SEER) database as a gold standard. In this way, we hope to better match procedures identified using Medicare claims data with SEER listed procedures. The focus of this research is observational (i.e. non-randomized) data. Well-run randomized clinical trials can provide the best level of evidence of treatment effects. However, randomized trials in the United States have suffered from poor accrual for many interventions. Despite the fact that well-designed randomized clinical trials should be the gold standard, well-designed observational studies might be the only method of obtaining inferences concerning comparative effectiveness for some cancer interventions. In cancer research, one of the most commonly used databases for observational research is the linked SEER-Medicare database. SEER-Medicare data has provided useful measurements of the effectiveness of a number of cancer therapies. Algorithms for identifying relevant treatment and diagnosis codes using Medicare data are often based on clinical reasoning and scientific evidence. One group of researchers, for example, developed an algorithm for identifying laparoscopic surgery among kidney cancer cases before claims codes for laparoscopic surgery were well developed. While such algorithms are useful for others pursuing similar investigations, there may still be substantial mismatch between treatment identified by the SEER cancer registry and treatment identified through Medicare claims. In this work, we propose developing a rigorous machine learning algorithm that can help researchers in better identifying treatments in Medicare claims data. Specifically, we will design a neural language modeling algorithm and implement a software system that finds vector representations of diagnosis and procedure codes. We plan on using the neural language modeling algorithm to learn vector representations from SEER- Medicare claims data where related procedure and diagnosis codes are "neighbors" (i.e. closely related). We will investigate whether the codes we identify within neighborhoods correspond to the procedure codes used for published SEER-Medicare studies. We will then design a software assistant interface that will allow an investigator to explore which codes are related to a given seed of diagnosis or procedure codes. Finally, we will investigate the sensitivity and specificity of the algorithm by comparing procedures identified using Medicare claims with procedures listed in the SEER database. We will replicate analyses from a published SEER-Medicare paper to investigate if estimated treatment effects differ when using our novel algorithm compared to using the algorithm in the published paper.