Using Computational Approaches to Optimize Asthma Care Management

摘要 这项研究将开发更准确的计算预测模型和一种新的自动解释 更好地识别可能从护理管理中受益最大的患者的功能。对于许多慢性病来说， 易损性高、疾病严重或护理障碍大的一小部分患者消耗最多 医疗资源和成本。为了改善结果和资源使用，许多医疗保健系统使用 预测性模型，用于前瞻性地识别高危患者并将他们纳入护理管理以实施 量身定制的护理计划。为了在服务能力有限的情况下从昂贵的护理管理中获得最大利益，只有患者 风险最高的人应该被录取。但是，目前的患者身份识别方法有两个局限性： 1)预测精度低会导致错误分类、浪费成本和次优护理。如果现有模型 如果用于护理管理分配，登记将错过50%的那些将受益最大的人 但包括其他不太可能受益的人。医疗保健系统通常没有足够的数据来进行模型培训和 许多患者的数据不完整。一个典型的模型只使用几个不良后果的风险因素，尽管 其中许多人都是众所周知的。此外，还没有发现许多关于患者和系统特征的预测变量。 2)没有解释预测的原因会导致对预测的采纳不佳和忙碌的护理 管理者会花费额外的时间，错过合适的干预措施。护理经理需要了解为什么 预计患者在分配到护理管理和形成量身定制的护理计划之前处于高风险状态。 现有的模型很少给出这样的解释，迫使护理经理进行详细的病历审查。 针对局限性，优化护理管理，为更多高危患者提供适当的 护理，这项研究将：a)提高通过计算识别高危患者的准确性，并评估潜力 对结果的影响；b)自动解释计算预测结果并评估对模型的影响 准确性和结果；c)评估自动解释对护理经理接受 预测和感知护理计划质量。使用案例将是影响9%美国人的哮喘，并导致 每年住院43.9万人次，急诊室就诊180万人次，费用560亿美元。哮喘专家和 计算机科学家将使用来自三个领先医疗保健系统的数据；一种新颖的、基于模型的迁移学习 不需要其他系统原始数据的技术；一种新颖的、基于模式的自动解释技术，它还 提高模型的通用性和准确性；新的数据源PreManage使患者数据更多 完整；以及关于患者和系统特征的新特征。这些技术可以促进临床 用于各种应用的机器学习，改进患者身份识别，并帮助形成量身定制的护理计划。焦点 将与临床医生进行小组讨论，以探索将该技术推广到慢性阻塞性肺疾病患者 阻塞性肺病、糖尿病和心脏病，这些疾病也需要护理管理。这个 结果将潜在地改变护理管理，以获得更好的结果和更有效的资源使用。

Abstract The study will develop more accurate, computational predictive models and a novel automatic explanation function to better identify patients likely to benefit most from care management. For many chronic diseases, a small portion of patients with high vulnerabilities, severe disease, or great barriers to care consume most healthcare resources and costs. To improve outcomes and resource use, many healthcare systems use predictive models to prospectively identify high-risk patients and enroll them in care management to implement tailored care plans. For maximal benefit from costly care management with limited service capacity, only patients at the highest risk should be enrolled. But, current patient identification approaches have two limitations: 1) Low prediction accuracy causes misclassification, wasted costs, and suboptimal care. If an existing model were used for care management allocation, enrollment would miss >50% of those who would benefit most but include others unlikely to benefit. A healthcare system often has insufficient data for model training and incomplete data on many patients. A typical model uses only a few risk factors for adverse outcomes, despite many being known. Also, many predictive variables on patient and system characteristics are not found yet. 2) No explanation of the reasons for a prediction causes poor adoption of the prediction and busy care managers to spend extra time and miss suitable interventions. Care managers need to understand why a patient is predicted to be at high risk before allocating to care management and forming a tailored care plan. Existing models rarely give such explanation, forcing care managers to do detailed patient chart reviews. To address the limitations and optimize care management for more high-risk patients to receive appropriate care, the study will: a) improve accuracy of computationally identifying high-risk patients and assess potential impact on outcomes; b) automate explanation of computational prediction results and assess impact on model accuracy and outcomes; c) assess automatic explanations' impact on care managers' acceptance of the predictions and perceived care plan quality. The use case will be asthma that affects 9% of Americans and incurs 439,000 hospitalizations, 1.8 million emergency room visits, and $56 billion in cost annually. Asthma experts and computer scientists will use data from three leading healthcare systems; a novel, model-based transfer learning technique needing no other system's raw data; a novel, pattern-based automatic explanation technique that also improves model generalizability and accuracy; a new data source PreManage to make patient data more complete; and novel features on patient and system characteristics. These techniques can advance clinical machine learning for various applications, improve patient identification, and help form tailored care plans. Focus groups will be conducted with clinicians to explore generalizing the techniques to patients with chronic obstructive pulmonary disease, diabetes, and heart diseases, on whom care management is also needed. The results will potentially transform care management for better outcomes and more efficient resource use.