On Friday 2nd September, I attended the short Post-Doc talks at Human Computer Interaction Institute of Carnegie Mellon University. This year was the 25th anniversary of the HCII Seminar Series as stated by Professor Myers. There were three 20-minute-long talks presented by the postdoctoral researchers. Since the talk of Dr Bae mostly interested me, I am presenting the short brief of her talk.
Using Passively Collected Sedentary Behavior to Predict Hospital Readmission
By Grace Sangwon Bae
Dr Bae presented her recent paper accepted for the proceedings of 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing (UbiComp 2016. Their study is mainly about the factors affecting the readmission of patients to hospitals after surgery. Readmitting patients after sugery is a significant concern for health-care providers. To support this, she presented some statistics. For example, 1 in every 7 post-surgery patients ar readmitted within 30 days of their discharge. In addition, readmissions cost about US$26 million yearly to Medicare. Readmissions are associated with high cost, seriously shorter suvival and stress of the patient and their family. Hence, if the factors affecting the readmissions are well-studied, they can be mostly prevented and their negative results can be mitigated.
Different than prior studies, they investigated the behavioral factors instead of medical predictors of readmission. Specifically, they considered sedentary behaviors (insufficient physical activity) of post-surgical cancer patients, such as daily step counts, sedentary bouts (duration of the time with no steps taken) and sedentary time (time in which step count is less than 100 in a minute). They used the passively collected sensor data from the Fitbit wearable devices worn by 25 post-operative cancer patients. They proposed two different models, which are step-only and behavior-based models. These two models are based on machine learning approach and they used Weka to run the algorithms. They trained Random Forest Classifier and used Information Gain and Gain Ratio algorithms to select redundant features. They compared the performance of two models in terms of prediction accuracy for readmission rates of the patients after the 5th, 10th and 15th days of their discharge. As the result, behavior-based model outperformed the step-only model. The most prominent limitation of their study was the small sample size (25 patients) and this might cause overfitting of the ML algorithms. As another limitation, they considered only behavioral factors, not the medical features such as syptoms of the patients. As the future direction, they point to inclusion of additional behavioral factors such as adherence to medications by the patient, patient’s diet and attendance to follow-up medications.
I found this talk the most interesting because it is about machine learning and sensory data in health-care domain, the areas which Iam also doing reasearch. The most interesting thing about the talk for me was that the results of such a study would go beyond being displayed on a conference proceedings to be read by researchers. The findings of the study would help the cancer patients to recover from post-surgery by stimulating them to force themselves to change their physical activity behaviors.