Machine Learning for Precision Treatments in Schizophrenia

项目摘要/摘要精神分裂症与精神病症状、情绪障碍、 认知缺陷、合并症、严重的社会和功能障碍，是 美国和世界各地的残疾问题。尽管抗精神病药物和心理社会治疗 对于精神分裂症的一些症状有效，但针对所有症状的有效治疗方案尚未建立。这个 治疗指南的主要局限性是依赖于测试有限治疗及其影响的随机对照试验 几乎没有症状和合并症。对所有方面的损害进行治疗的试验是 复杂得令人望而却步。数据驱动的机器学习(ML)可以使用大型观测数据来解决这一差距 包含现实世界实践中使用的复杂和有效的养生方法的信息的数据集。ML罐头集群 具有共同特征的个体，并确定针对他们的精神疾病和 临床合并症。这些新的治疗方案是可能的精准治疗。ML算法可以 然后预测以患者为中心的这些不同集群(或类别)的关键结果 治疗方案。检查这些预测危重病例的治疗方案的比较效果 结果是下一步至关重要的一步。独特的药物流行病学方法和观察数据可以 模拟临床试验。倾向性评分方法解决了令人困惑的问题，模仿了通过 随机对照试验中的随机化。这些工具将决定哪种精确治疗方案是最有效的 用于这些数据集中的类。ML结果的相关性取决于数据质量。索赔金额最大的， 大多数具有全国代表性的样本反映了现实世界的社区实践模式，但使用了计费代码 最初不是为研究而设计的。电子健康记录(EHR)很广泛，但由于偏颇于 由于详细程度的原因，不完整的记录具有不确定的准确性和复杂性。这项建议 我将通过对样本进行ML分析来确定这些数据集类型的优势和局限性 数据集，医疗补助分析摘录(MAX)全国样本，以及观察性健康数据科学和 信息学(OHDSI)网络纽约长老会医院(INYP)EHR。对该项目的改进将 将更传统的多变量和回归技术与ML结果进行比较，以确定ML是否 提供了更多信息。为了解决“研究-实践”的差距，ML的结果将被转化为 个性化治疗规则，为精神分裂症治疗的临床实践提供参考。在接受培训后 非监督和监督学习在培训目标A和B中，研究目标1将确定班级及其 数据集和研究目标2中实施的治疗将预测这些治疗的结果： 急诊科就诊、再次入院时间和合并症发生率。研究目标3将使用 在培训目标C中学习的药物流行病学方法以比较治疗的有效性， 支持在本K奖结束时提交的R01，以在国际EHR数据集中测试有效性。

Project Summary/Abstract Schizophrenia is associated with psychotic symptoms, mood disturbances, deficits in cognition, comorbidities, significant social and functional impairment and is a leading cause of disability in the U.S. and worldwide. Although antipsychotic medications and psychosocial treatments are effective for some symptoms of schizophrenia, effective regimens for all symptoms are not established. The primary limitation of treatment guidelines is reliance on RCTs that test limited treatments and their effects on few symptoms and comorbidities. Trials of treatments administered to address all aspects of impairment is prohibitively complex. Data driven machine learning (ML) can address this gap using large observational datasets with information about complex and effective regimens used in real-world practice. ML can cluster individuals with shared characteristics and identify unique regimens administered for their psychiatric and clinical comorbidities. These new treatment regimens are possible precision treatments. ML algorithms can then predict critical patient-centered outcomes for these different clusters (or classes) administered these treatment regimens. Examining the comparative effectiveness of these treatment regimens that predict critical outcomes is an essential next step. Unique pharmacoepidemiologic methods with observational data can simulate clinical trials. Propensity score methods address confounding, mimicking balance achieved by randomization in RCTs. These tools will determine which precision treatment regimens are the most effective for the classes in these datasets. Relevance of ML findings depends on data quality. Claims have the largest, most nationally representative samples reflecting real-world community practice patterns but use billing codes not originally designed for research. Electronic health records (EHR) are extensive but limited due to bias from incomplete records with uncertain accuracy and complexity due to their granular level of detail. This proposal will establish the strengths and limitations of these dataset types by conducting ML analyses on exemplar datasets, a Medicaid Analytic eXtract (MAX) national sample, and the Observational Health Data Sciences and Informatics (OHDSI) network New York-Presbyterian Hospital (iNYP) EHR. An enhancement to this project will compare more traditional multivariate and regression techniques to the ML findings identifying whether ML provides additional information. To address the “research-practice” gap the ML results will be translated into personalized treatment rules to inform clinical practice for schizophrenia treatment. After training in unsupervised and supervised learning in Training Aims A and B, Research Aim 1 will identify classes and their administered treatments in the datasets and Research Aim 2 will predict outcomes of those treatments: time to emergency department visit, time to re-admission and incidence of comorbidities. Research Aim 3 will use pharmacoepidemiologic methods learned in Training Aim C to compare effectiveness of the treatments, supporting an R01 submitted at the end of this K-award to test effectiveness in an international EHR dataset.