(557.6) Machine Learning to Identify Regional and Segmental Dysfunction during Type 2 Diabetes Mellitus Progression

Poster Board Number: E129

Saira Rizwan (West Virginia University), Andrya Durr (West Virginia University), Anna Korol (West Virginia University), Quincy Hathaway (West Virginia University), Amina Kunovac Hathaway (West Virginia University), Andrew Taylor (West Virginia University), Mark Pinti (West Virginia University), John Hollander (West Virginia University)

Diabetes mellitus results in numerous co-morbidities, the most serious of which is cardiovascular disease (CVD). Speckle tracking strain-based echocardiography (STE) can be used to assess regional and segmental dysfunction prior to the onset of clinically recognizable symptoms, providing a unique opportunity to assess localized patterns of cardiac dysfunction. At present, STE has been minimally used to assess localized impacts of CVD in type 2 diabetes mellitus (T2DM). Therefore, the aim of this study was to utilize machine learning (ML) to determine the ability of STE parameters to predict T2DM and to elucidate regional and segmental impacts in a temporal fashion. Echocardiography data were collected in wild-type and Db/Db mice at 5, 12, 20, and 25 weeks. Seventeen parameters were ­used for analyses: Complete, pulse-wave Doppler, M-mode, Global, Segmental, AntFree, PosteriorFree, Anterior, Posterior, Septal, Free, LatWall, PostWall, InfFreeWall, PostSeptal, and AntSeptum. Support vector machine analysis determined that STE parameters were the most accurate predictors of T2DM. Further, ReliefF feature selection was used to evaluate prevalence of each region and segment. We determined the Anterior region and AntSeptum segment to be the most prevalent (75% of the time), at weeks 5, 20, and 25 (82%, 93%, and 91% testing accuracies, respectively). The Septal region was shown to be most predictive (75% of the time), at weeks 5, 20, and 25 (84%, 93%, and 98%, respectively). Further, the InfFreeWall (76%) and LatWall segments (98%) were most predictive at 5 and 12 weeks, whereas the AntSeptum and AntFree segments (96% each) were most predictive at 20 and 25 weeks, respectively. Specifically, in some cases such as 25 weeks, the Septal region and the AntFree segment present identically, but the AntFree segment may provide a more focused target. In summary, these data demonstrate that patterns of regional and segmental dysfunction exist in the T2DM heart and can be observed in a spatiotemporal fashion using ML.

This research was supported by the NIH NHLB grant HL128485 and the WVU CTSI grant U54GM104942 awarded to JMH; the NSF IGERT- Ramp;amp;E in Nanotoxicology at WVU Fellowship grant 1144676 and AHA Predoctoral Fellowship (AHA 17PRE33660333) awarded to QAH; an AHA Predoctoral Fellowship (AHA 20PRE35080170) awarded to AK; WV-INBRE funded by NIH Grant (P20GM103434), and the Community Foundation for the Ohio Valley Whipkey Trust awarded to JMH. Imaging experiments and image analysis were performed in the WVU Animal Models amp;amp; Imaging Facility supported by the WVU Cancer Institute and NIH grants P20 RR016440, P30 RR032138/GM103488, and S10 RR026378.