Analyzing Streaming Multi-Sensor Data to Predict Stroke in Preterm Babies

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT In this Phase I SBIR application for Analyzing Streaming Multi-Sensor Data for Predicting Stroke in Preterm Infants, we propose developing statistical software for predicting adverse medical events using sensor data from preterm infants. While medical sensor data is becoming widely available as part of the Internet of Medical Things (IoMT), healthcare provider’s ability to use these data is limited by a lack of real-time predictive algorithms for detecting deteriorating conditions in patients. Very low birth-weight preterm infants have a high risk of experiencing intraventricular hemorrhage (IVH), a serious form of bleeding in the brain associated with high rates of mortality and other serious conditions such as cerebral palsy. The algorithm we will develop uses an innovative approach of transforming sensor data into graphs of associations and applies decision rules from statistical process control to determine when a patient’s data indicates an adverse medical event such as an IVH. If successful, this algorithm can be implemented in neonatal intensive care units (NICU) to provide real-time alerts to hospital staff, allowing for early detection and treatment of IVH before it causes severe damage. Two aims are proposed: to develop the software and test it on an existing, curated, large retrospective cohort of NICU data collected at Washington University (Aim 1); and to compare the accuracy of the method and software to existing predictive models of neonatal IVH and other outcomes (Aim 2). The first aim builds upon existing proprietary software for object oriented data analysis and encompasses testing different methods of measuring and relating sensor data, as well as evaluating decision rules for the graphical objects created from these data. The second aim involves testing the accuracy and specificity of the alerts created by this method to ensure it can detect adverse events significantly better than chance or existing algorithms, and to ensure it does not substantially contribute to the problem of false alerts. If successful, this project will lead to a Phase II proposal to test the algorithm in real-time inside an NICU with nursing staff and develop the algorithm into a marketable software platform. Phase II would also involve extending the testing of this software for other types of sensor data and medical events, such as monitoring medical conditions for adult patients or nursing homes, etc. This project has commercialization potential both in providing an important tool for improving patient care in NICUs, and in the broader context of developing tools for predicting adverse medical events from all types of IoMT data.