Time-dependent Survival Neural Networks for Predicting Incoming Workload and Order Turn Around Time in a Radiology Service

Intelerad是一家加拿大公司，总部设在蒙特利尔，在卡尔加里和多伦多设有办事处。该公司设计和开发的医疗成像软件解决方案，从其加拿大办事处，使医院，成像中心，放射阅读组访问和解释医学图像，以建立诊断。 在这个项目中，我们建议调查工作量和订单周转时间（达特）预测问题，这是Intelerad千里眼预测分析的核心，Intelerad正在构建的仪表板解决方案，以提供对放射科实践操作的深入了解。我们将探索，调整和测试我们最近开发的事件发生时间模型，特别是时间依赖生存神经网络（TSNN）模型，以实现准确的预后。我们将研究用于预测一天中不同时段的多个神经网络的设计。这包括设计1）一个新的学习目标函数来描述数据流上事件的风险和一个评估度量来估计TSNN的时间相关输出的误差：2）综合的网络结构来考虑TSNN构造中的输入工作量;和3）用于发现工作流程中成像模态和阅读组之间的交互模式的流的排序表示，以便识别异常事件并提高预测准确性。该项目将为HQP培训提供绝佳机会，因为他们将能够直接与我们的行业合作伙伴就真实的数据进行合作。该项目的产出将使我们的合作伙伴受益，扩大其现有仪表板的范围，以及如何为复杂的工作负载和达特分析开发未来产品的内部知识。从图像采集到结果可用性的改进达特，以及更好的护理质量，将使加拿大的医疗保健提供者能够根据预期的工作量更好地为专科医生配备人员。

Intelerad is a Canadian company with its Head Office in Montreal and additional Canadian offices in Calgary and Toronto. The company designs and develops medical imaging software solutions from its Canadian offices which allow hospitals, imaging centers, and radiology reading groups to access and interpret medical images in order to establish a diagnosis. In this project, we propose to investigate the workload and order Turn Around Time (TAT) prediction problem which is at the core of Intelerad Clairvoyance Predictive Workload Analytics, a dashboard solution that Intelerad is building in order to provide deep insights into the practice operations of a radiology department. We will explore, adapt and test the time-to-event models we have developed recently, especially a time-dependent survival neural networks (TSNN) model to achieve an accurate prognostic. We will look at the design of multiple neural networks dedicated to the prediction of different periods of a day. This includes design of 1) a new learning objective function to describe the risk of an event over a data stream and an evaluation metric to estimate the error of TSNN's time-dependent outputs; 2) comprehensive network structures to take account of incoming workload in TSNN construction; and 3) a sequencing representation of streams for the discovery of interaction patterns between imaging modalities and reading groups in the workflow in order to identify anomaly events and improve predictive accuracy.This project will provide an excellent opportunity for HQP training as they will be able to work directly with our industrial partners on real data. The output of this project will benefit our partner by expanding the scope of its existing dashboard and its internal knowledge of how to develop future products for complex workload and TAT analysis. Improved TAT from image acquisition to results availability, as well as a better quality of care, will allow Canadian healthcare providers to better staff subspecialists based on expected workloads.