SPUR 2019: A Temporal Analytics Framework for Predicting Early the Co-occurrence of Shocks in the Intensive Care Unit


Several types of patients are hospitalized in intensive care units (ICU), such as patients with acute myocardial infarction, bleeding, and sepsis-yielding shock, a life-threatening condition, and, in many cases, the co-occurrence of shocks or death. One-third of ICU patients suffer from some types of shocks. Of them, cardiogenic shock comprises approximately 20%, hypovolemic about 20%, and septic shock about 60% of cases. Shock types may have temporal co-occurrences that lead to further complications including mortality. Therefore, clinicians may depend on their intuition when analyzing time-sensitive information to make several clinical decisions for predicting early the likelihood of more than one type of shock at occurring at the same time (i.e., multi-shock). Most previous modeling studies have focused on predicting a single shock rather than multi-shocks. Moreover, most of them use machine learning techniques in their clinics that may not accurately predict such temporal co-occurrences among different types of shocks. The Abdelrahman lab has developed many temporal solutions that have been used effectively in critical care settings. Our primary goal is to develop a novel temporal solution that leverages our prior preliminary results, MIMIC dataset, and advanced machine learning techniques to better predict patient outcome in the shock domain. The proposed solution impact is to support the clinicians with early and accurate predictions of multi-shocks. This project entails collaborations between informatics researchers and clinical experts to identify challenges and propose relevant solutions and will require the student selected for the project and lab students to exchange ideas and solve problems together.

This SPUR project is funded by a supplement to the National Library of Medicine Training grant T15LM007124-22; Wendy Chapman (PI), Julio Facelli (co-I).

Student Role

The student selected in the project will use MIMIC-III (Medical Information Mart for Intensive Care III) patient demographic data, symptoms, signs, and repeated measure timestamp data, such as vital signs, laboratory tests, procedures codes, and medications. The student will review the literature to identify how to use these factors to predict early each type of shock separately or jointly. She or he will use queries to pull the necessary data from MIMIC. The student will master concepts of machine learning techniques through sklearn and tensorflow packages and will use these concepts and the packages to conduct data cleaning methods and descriptive analysis to filter the noise and non-risky factors, respectively, from the data. Also, she or he will run machine learning techniques currently used in clinics and our lab proposed solution on the pulled data. The student then will prepare a report to compare the results among different solutions that will require her or him to draft a publication including literature review, exploratory analysis, results, discussion, limitations, and future work. If the student lacks the expertise to complete the mentioned tasks, the lab will support her or him with the necessary documents, resources, and guides. Dr. Abdelrahman will mentor the project student and the lab students and encourage the lab students to communicate and exchange ideas to help the project student acquire the necessary expertise. The project student will communicate with the lab students and will attend meetings with other lab members and project partners. The student will present her or his findings from the meetings and the project tasks on a regular basis to get feedback from the lab members.

Student Learning Outcomes & Benefits

The student selected in the project will learn the perspectives of the clinical domain, databases, machine learning, and temporal modeling. From a clinical domain perspective, the student will gain insights into (i) types of shock, (ii) characteristics of each type and complications (iii) predictors of each type, (iv) why the co-occurrence between types happens, and (iv) the interpretations of predictor temporality behaviors and prediction results. From a database perspective, the student will understand the database structure and will run lab queries by which she or he will gain a better understanding of the database-structured query language. From a machine learning perspective, she or he will understand the concepts and technical aspects of sklearn and tensorflow packages, which will then enable her or him to run them in the project. From a temporal modeling perspective, the student will understand: (i) the difference between modeling temporal and static events and (ii) the extraction and analysis of the temporality events. Of note, if the student has this expertise before joining the lab, she or he will gain more practice and hands-on skills to help propose a new solution for the project questions. On the other hand, if the student does not have enough knowledge or skills to gain such an understanding, Dr. Abdelrahman will either sharpen the student’s knowledge or adapt/minimize the outcomes of each project perspective separately for her or his maximum benefits from the project. The student will acquire presentation, communication, collaboration, and publication skills and will be trained to work independently and in-group. These skills will support her or him in future data science careers in industry or academia.

Samir Abdelrahman
Assistant Professor

Biomedical Informatics
School of Medicine

Dr. Abdelrahman has had more than a decade of mentoring approximately twenty-five undergraduate and fifteen graduate students. His goal is to gradually educate students about the basic concepts such that, eventually, they become completely independent in developing and validating their ideas.. Therefore, he starts every project by setting up clear milestones with clear definitions of goals, input, and output. He meets with the involved students at the time of the first milestone to let them know the project details/team and to discuss their opinions about the next milestones. Based on these discussions, Dr. Abdelrahman sets up the time-intervals for the project’s regular meetings, informs the students how they should report their findings and communicate with each other, and suggests modifications of the milestone(s), if needed. For the next milestones, he assesses the students’ performances and identifies any challenges that need to be addressed. He also encourages the students who have successfully solved their problems to develop new ideas. For those who have challenges, he meets with each student individually to understand her or his problems and to find an adequate solution. If the solution requires any change in the student’s activities, Dr. Abdelrahman adapts the milestones to reflect the student’s skills and thoughts. As a mentor, he also encourages his lab students to share ideas, and he organizes social meetings so the students can interact outside the lab environment.