Processing of temporal medical data for identifying high-risk ICU patients in hospital using time-aware machine learning models
- 0 Collaborators
Medical history data is a currently unused treasure, which could be used to automate workflow in the hospital. This proposal is believed to help doctors in decision-making, disease surveillance and personalize medical services, which could make health services more accurate and precise as well as reducing healthcare costs. Today, my project is to develop an algorithm for identifying high-risk ICU patients in the hospital, which later could be used as an app. To reach my goal, I need to extract data from a medical database (MIMIC III, which is a critical care medical database and exactly meet my requirements.). Then for ICU data, new time-aware machine learning models could be used for continuous lab test or vital signs data in order to mine the pattern for this time-series data. ...learn more
Project status: Under Development
Overview / Usage
Inpatient mortality is a kind of adverse events in hospital, which is one of the worst clinical outcomes. To avoid certain fatal adverse events, an alarm system is very necessary to identify some high-risk patient especially in ICU, which is a site doctors should pay more attention to. The advanced intervention could be performed to greatly reduce some avoidable adverse events or deaths. In traditional alarm systems, the decision of whether a patient is urgent and need additional interventions or not is according to the doctor’s experience, which is sometimes not very accurate and may miss the best intervention plan. In the big data area, machine learning methods could be applied to identify the pattern of high-risk patients by mining the Electronic Health Records.
Methodology / Approach
EHR (Electronic Health Record) is a digitalized patient health information chart, which contains different categories of data, including lab and vital signs, which matter for ICU patients. For those critical patients especially the ICU patients, lab and vital signs data are captured several times per hour. The recorded data are temporal, continuous. However, they also are irregular and inconsistent. Many studies opt for using smaller summary statistics such as average or standard for these kinds of irregular and temporal data. Few studies use longitudinal measurements for patients. Nevertheless, I think this should be a good opportunity to take the data changes from independent patients into consideration and it will make results that are more valuable from the temporal data analysis.
In my project, the machine learning method will be applied to temporal data analysis for this real-world problem. I will try different time-aware modeling techniques
Technologies Used
MIMIC-III, an accessible critical care database(https://mimic.physionet.org/) could be used for this projects. It contains about 50k ICU admission records, including vital and lab test data from ICU critical care which fit our requirements well.