Use of Freetext Clinical Reports for Prediction of 30-Day Psychiatric Readmissions / Deep Neural Networks for Classification of Dysplastic Structures in Neonatal Brain MRI

Seminar Date: 
2016-04-15
Seminar Time: 
11am - 12pm
Seminar Location: 
5607 Baum Boulevard, Room 407A
Presenter: 
Jose Posada and Rafael Ceschin

Posada:  According to the substance abuse and mental health services administration (SAMHSA) 43.1 million of adults (18.1%) in the US experienced some form of mental illness. Hospitalization for these of diseases is increasing at a faster rate than any other type of hospitalization. From the psychiatric hospitalized patients, 15% of all discharges were readmitted within 30 days and the costs among those readmissions are higher than for any other readmission cause. Moreover, CMS releases individual hospital readmission rates to the public as an indicator of quality of care. One of the key steps to reduce 30-day readmission rates is to identify patients at high risk of 30-day readmission. However, existing studies including our previous one using only structured psychiatric data such as demographics In this study, we hypothesize we can better predict 30-day psychiatric readmissions from unstructured clinical reports.

Ceschin:

Neonates born with congenital heart defects (CHD) often present with associated neurological deficits due to both environmental and genetic factors. Of particular concern is the maldevelopment of key brain structures, which can lead to poor neurocognitive outcomes. Currently, full classification of these structural deficits is very labor intensive, requiring a trained neuro-radiologist to carefully examine multiple structures using a detailed protocol involving both volumetric and morphological observations. Our goal is to automate this process to the best of our capabilities.

We have developed a robust structure segmentation pipeline, which parcellates the neonatal brain into over 50 structures of interest and extracts their 3-dimensional shape and volume. We can use the output of this pipeline as the input to a machine learning algorithm capable of predicting each structure’s dysplastic measurement based on both shape and volume. Here, I propose the use of deep neural networks, in particular Convolutional Neural Networks (CNNs), to predict each structure’s likelihood of dysplasia. CNNs have proven to be highly effective tools in the domain of computer vision, making them a strong candidate for our pipeline.

^