Using AI for Predicting Syncope in Older Persons Nandu Goswami - - PowerPoint PPT Presentation

1

Using AI for Predicting Syncope in Older Persons

Nandu Goswami

Chair of Physiology Division Otto Loewi Research Center of Vascular Biology, Immunity and Infmammation Medical University of Graz, Austria Director of Research of Health Sciences, Physiotherapy and Social Gerontology Alma Mater Europea Maribor, Slovenia Co-ordinator of Falls Prevention Task Force European Innovative Partnership Active & Healthy Aging

2

2013

2060

17 %

30 %

EU Aging Report, Brussels

Countries Costs

EU Aging Report, Brussels

3

4

• 65+ year old patients  40 % acute hospitalizations
• Poor outcomes:

… high 1 year mortality … 30 % functional decline … high re-admission rates … higher home healthcare usage

5

Immobilizatio n De-conditioning Falls / Fear of falling further

6

• Keeping ambulatory persons mobile
• Getting bed-confjned persons re-

mobilized

Singh et al. (2008). Mayo Clinic Proceedings, 83(10), 1146-1153.

7

8

Grasser E, Goswami N, Hinghofer-Szalkay H. Presyncopal cardiac contractility and autonomic activity in young healthy males. Physiol Res 2009; 58, 817-26

9

10

• Hardware:
• 16 nodes of Prometheus were used :
• K40 nvidia gpus,
• Approx. 12 days of computing
• mostly for setting right parameters
• Deep learning methods:
• LSTM (Long short-term memory): Better

than the classic statistical methods (ARIMA, ES1, ES2, Winter-holds, moving-average)

11

• 85 % accuracy of classifjcation with LSTM
• Heart rate (HR) and mean blood pressure

used

12

13

• More detailed presentation on what was done
• More time series to be taken
• Stroke volume & stroke index (accounts for sex and

age)

• Total Peripheral Resistance changes accordingly to

mBP

• Incorporate more information
• Gender Predisposition
• Age Predisposition

15

• Model trained on syncope dataset
• Dataset was balanced by sub-sampling regular data so it

matched the number of syncope data

• All difgerent classes of Syncope were taken
• Threshold was set of 0.7 (assumed that a patient would faint)
• Reduced signal of syncope was taken

○ 500 points were truncated the beginning ○ 500 points removed at the end