Is That A Sugar High or High Sugar? Ask The Algorithm!
Friday, February 16, 2018
10:30 AM – noon
Speakman Hall Suite 200
We use methods based on machine learning to predict the transition of health state in chronic care patients. We collected data from a large, tertiary care, multi-specialty urban hospital system. In these hospitals about 2450 patients were in a diabetes case management program and from this cohort we obtained the EMRs of 1,687 patients. We use a two-stage machine learning algorithm that combines structured and unstructured data to predict the transitions in patients’ health states. We use a technique that we term ‘Differential Parsing’ where the output of the algorithm’s first stage prediction is used to identify those instances where there are key information signals in unstructured data. We then use LDA algorithm to extract decision rules from unstructured data which then enter as inputs for the second stage of the prediction. We construct four ML algorithms and contrast these with regression-based models. We find that the use of unstructured data improves the prediction accuracy of ML algorithms and they outperform regression-based models.