Track 6. Technological innovations in biomedical training and practice
Authors: Elena Martín-González, Rodrigo de-Luis-García, Pablo Casaseca-de-la-Higuera, José Ramón Garmendia-Leiza, Jesús Andrés-de-Llano and Carlos Alberola-López
Mapping Raw Acceleration Data on ActiGraph Counts: A Machine Learning Approach
1. ELENA MARTÍN-GONZÁLEZ,UNIVERSIDAD DEVALLADOLID
RODRIGO DE-LUIS-GARCÍA,UNIVERSIDAD DEVALLADOLID
J.P. CASASECA-DE-LA-HIGUERA, UNIVERSIDAD DEVALLADOLID
J.R. GARMENDIA-LEIZA, COMPLEJOASISTENCIAL UNIVERSITARIO DE PALENCIA
J.ANDRÉS-DE-LLANO, COMPLEJOASISTENCIAL UNIVERSITARIO DE PALENCIA
CARLOS ALBEROLA-LÓPEZ, UNIVERSIDAD DEVALLADOLID
MAPPING RAW ACCELERATION DATA ON ACTIGRAPH
COUNTS:A MACHINE LEARNING APPROACH
Salamanca, October 2018
2. Actimetry is a valuable tool for the objective diagnosis and the monitoring of different pathologies
There are different types of actimeters in the market
26/10/2018 2ELENA MARTÍN GONZÁLEZ | TEEM 2018, SALAMANCA
A mapping must be found between data recorded with an approved clinical diagnostic actimeter
(counts) and acceleration data recorded with a commercial actimeter
Goal
Map raw acceleration data from a commercial actimeter
on ActiGraph counts, using a machine learning approach.
Is it possible to use commercial devices, not approved for medical
use, to obtain information that has some clinical relevance ?
MOTIVATION
3. 12 healthy volunteers (5 women, 7 men; 15-56 years)
Both actimeters on the wrist of their non-dominant hand
Records between 13 and 83 min (walking and spontaneous physical activity)
MSB data: in-house mobile application
Raw data: 31.25Hz
ACT data:Actilife software
Raw data: 30Hz
Counts: 1s
26/10/2018 3ELENA MARTÍN GONZÁLEZ | TEEM 2018, SALAMANCA
MATERIALS
ActiGraph wGT3X-BT Microsoft Band 2 Smartband
4. 26/10/2018 4ELENA MARTÍN GONZÁLEZ | TEEM 2018, SALAMANCA
Step 1. Preprocessing: resampling and alignment
Step 2. Correspondence between MSB raw data and ACT counts
Step 3.Approximation machine learning neural networks
Train different architectures of neural networks
METHODS
CONFIGURATIONS
One hidden layer (7 networks)
[10 – 70; 10] neurons
Two hidden layers (16 networks)
[10 – 40; 10] neurons
[2 – 8; 2] neurons
1
2
1
Scheme
ො𝑥 ≃ 𝑔(𝒚)
5. RESULTS (I)
26/10/2018 5ELENA MARTÍN GONZÁLEZ | TEEM 2018, SALAMANCA
Neural network with 1 hidden layer
10 neurons
Correlation coefficient: 0.893005
Slope: 45.01º
95.13% of the differences are within the 95
per cent confidence interval
6. RESULTS (II)
26/10/2018 6ELENA MARTÍN GONZÁLEZ | TEEM 2018, SALAMANCA
Correlation coefficient: 0.899548
Slope: 44.98º
95.11% of the differences are within the 95
per cent confidence interval
Neural network with 2 hidden layers
10, 6 neurons
7. It is possible to approximate the function that provides counts from the MSB acceleration RD
Any actimeter available in the market (even if it does not have validity in clinical diagnosis).
Use of these actimeters in studies of the matter, given that these devices are available on the market to
the general public, not only limited to physicians and researchers
Limitations:
The database of MSB records has a small amount of subjects
The brevity of the records
Future work:
Extend the database
Test other machine learning, such as random forests or deep learning
26/10/2018 7ELENA MARTÍN GONZÁLEZ | TEEM 2018, SALAMANCA
CONCLUSIONS
8. Simultaneous records from ACT and MSB actimeters
Preprocessing
Correspondence between MSB raw data and ACT counts
Function approximation machine learning neural networks
Train
Test
We can approximate the function that provides counts from the MSB acceleration RD
This fact favors the use of these actimeters in studies of the matter, given that these devices are available
on the market to the general public, not only limited to physicians and researchers
26/10/2018 8ELENA MARTÍN GONZÁLEZ | TEEM 2018, SALAMANCA
SUMMARY
Goal
Map raw acceleration data from a commercial actimeter
on ActiGraph counts, using a machine learning approach.
9. ELENA MARTÍN-GONZÁLEZ, UNIVERSIDAD DEVALLADOLID
emargon@lpi.tel.uva.es
THANKYOU FOR LISTENING
MAPPING RAW ACCELERATION DATA ON ACTIGRAPH
COUNTS:A MACHINE LEARNING APPROACH
Salamanca, October 2018