Massachusetts General Hospital
Neurology
Sudeshna Das, Ph.D. is an Assistant Professor in Neurology at Massachusetts General Hospital (MGH) and Harvard Medical School (HMS). She directs the Data Core of the Massachusetts Alzheimer’s Disease Research Center (MADRC) and the MGH Biomedical Informatics Core. Her research focuses on applying statistical inference and machine learning to high-dimensional molecular -omics datasets and healthcare big-data to advance brain research in health and disease. She develops and applies deep learning algorithms to electronic health records for Alzheimer’s disease research and uses causal-inferencing techniques to emulate in-silico trials for drug repurposing for dementia. She collaborates closely with physician-scientists to understand the mechanistic basis of Alzheimer’s and other neurodegenerative disorders.

Moderator of 1 Session

Session Type
SYMPOSIUM
Date
Fri, 18.03.2022
Session Time
02:45 PM - 04:45 PM
Room
ONSITE: 133-134

Presenter of 1 Presentation

ACTIVE DEEP LEARNING TO DETECT COGNITIVE IMPAIRMENT IN ELECTRONIC HEALTH RECORDS

Session Type
SYMPOSIUM
Date
Fri, 18.03.2022
Session Time
02:45 PM - 04:45 PM
Room
ONSITE: 133-134
Lecture Time
03:00 PM - 03:15 PM

Abstract

Aims

Dementia is under-recognized in the community, under-diagnosed by healthcare professionals, and under-coded in claims data. Timely diagnosis of dementia, however, is important for patients and their caregivers to plan for the future. We aim to screen electronic health records (EHR) for signs of cognitive impairment.

Methods

To identify patients with cognitive impairment in EHR, we applied a deep learning based natural language processing (NLP) algorithm to unstructured clinician notes and compared the model’s performance to a baseline model that used regularized logistic regression with structured data (dementia related diagnosis codes and medication). We trained and evaluated the algorithm with a seed set of patients with detailed chart review and adjudication of cognitive status by a team of experts. Next, we used an active learning loop to continually improve model performance and select candidate cases for labeling using a combination of uncertainty and diversity measures.

Results

The baseline model with diagnosis codes and medications trained on the initial seed set (N=943) had an area under the receiver operating characteristic curve (AUROC) of 0.79. The deep learning model improved the AUROC to 0.92 and increased sensitivity of dementia detection from 0.59 to 0.79, at similar level of specificity. Of the patients which were predicted positive, 16.1% did not have a dementia-related diagnosis code suggesting that many patients with cognitive impairment may be undiagnosed.

Conclusions

Automatic processing of EHR with deep learning tools can be used to screen for patients with cognitive impairment who could benefit from a cognitive evaluation or be referred to specialist care.

Hide