Premium
[TD‐P‐003]: THE NATIONAL ALZHEIMER'S COORDINATING CENTER: QUERYING THE DATABASE AND REQUESTING DATA
Author(s) -
SchwabeFry Kristen,
Bollenbeck Mark,
Teylan Merilee,
Beekly Duane,
Thomas George,
Hubbard Janene,
Jacka Mary,
Wu Joylee,
Besser Lilah M.,
Kukull Walter A.
Publication year - 2017
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1016/j.jalz.2017.06.2599
Subject(s) - computer science , database , information retrieval , table (database) , sql , download , world wide web
hemisphere findings, we evaluated the diagnostic utility in adding markers of hemispatial neglect to our previous baseline algorithm. Methods: We used the publicly available DementiaBank dataset (257 interviews from 169 AD patients and 242 interviews with 99 healthy controls). In addition to our baseline algorithm, which included 353 lexical and acoustic markers, we evaluated three approaches to dividing the Cookie Theft image: Halves, strips and quadrants, as seen in figure 1. For each given division, we compiled a list of information units (info-units) that are contained in each region (e.g. the info-units “stool” and “mother” are contained by the left and right halves, respectively). For each region we then recorded four measures from a given transcript: 1) Number of infounits mentioned 2) ratio of info-units to all words 3) ratio of unique info-units to all possible info-units in the region 4) ratio of unique info-units to total mentioned info-units (a measure of redundancy). We also included quadratic interaction terms between regions. We then performed a 10-fold cross validation procedure with a correlation-based feature selection preprocessing phase and trained a logistic regression model using each of the halves, strips, and quadrants approach, and compared against baseline. Results: The halves model [PPV 0.84, 95%CI 0.80-0.86, NPV 0.81(0.74 – 0.88)] and strips model [PPV 0.84 (0.77 – 0.91), NPV 0.82 (0.76 – 0.88)] but not the quadrants model [PPV 0.81 (0.74 – 0.87), NPV 0.81 (0.75 – 0.87)] showed a trend towards improvement from baseline. Conclusions: Including markers of hemispatial neglect to a machine learning algorithm analyzing lexical and acoustic speech features may improve diagnostic accuracy of AD versus healthy controls.