Injuries in sport kost milions. Can we prevent it? For the bi-monthly Twente Data Meetup, Tatiana Goering-Zaburnenko gave a presentation on sport and health related projects, carried out in the research group Ambient Intelligence.
Twente Data Meetup - Data in sport and what can we do with it (Tatiana Goering-Zaburnenko, Ambient Intelligence, Saxion)
1. Data in Sport and
What we can do with it
Tatiana Goering-Zaburnenko
Data MeetUp, Saxion Enschede
7 November 2019
2. Overview
• Project problem definition
• Approach
• Results till now
• Other health related projects from AMI
3. Background
Researcher &
Project leader
Ambient Intelligence
Saxion, Enschede
PhD in Informatics
University of Twente,
Enschede
MSc in Math
& Applied Math
Yaroslavl, Russia
Applied Data Science
5. Many parameters are tracked in sports
(HR, speed (max, average,..), accelerations/ decelerations,
sprint distance, total distance etc.)
Can we use this information to predict injuries?
Previous project: Tried to predict injury
6. • Insufficient registered injuries to make robust
conclusions
• Software was made specifically for one team,
one measuring device (and didn’t allow to
experiment with different input / output parameters)
Problems
7. Focus on Load Monitor
try to predict internal player load based on
external training load
optimal training / recovery balance
reduction of injuries
Current project
8. Consortium partners (14)
• Universities of Applied Science (2):
Saxion (AmI, Health & Wellbeing),
Hanze
• University: UMCG (University
Medical Center Groningen)
• Practice partners (5): football (4),
volleyball (1)
• Physiotherapy (3)
• Roessingh R&D
• 360SI
• CE-Mate / Medimate
9. When is the training
successful?
GOAL = maximum training effect
WITHOUT injuries
Supercompensation
high
external load
short-term
loss of
fitness
fitness
improvement
10. Internal* player
load =
reaction of the body
to the external load
External player load
=
the work completed by the
athlete, measured
independently of his or her
internal characteristics
Measuring physical
condition
Player load
*Subjective and objective
11. Load Monitor
• External load parameters
(player load, actions p / m,
high intensity distance,
distance, ..)
• Internal load parameters
(HR, HR exertion index, time
in high HR zones, ..)
• Subjective parameters (RPE,
TQR, ..)
Load
Monitor
Training data from
the past
(personalized)
Expected current
internal load
Actual measured
internal load
Conclusion
Physiological model ML model(s) Validation
& Software
12. 1. Differences per association
1. Different measuring instruments / garments (Polar, Catapult, Johan Sports,
etc.)
different variables (speed, speed above 15 /18 km/h, HR, HRzone, etc.)
2. Different accents in the model (what a club wants to predict: training effect,
performance, etc.)
2. Input parameters: which are the most relevant?
3. Output parameters
4. Choice of ML algorithm(s)
5. Validation
Sub questions:
• What are the wishes of clubs regarding visualization? (user interface)
• How to make a (simple) system that allows a trainer to experiment with models
(“play” with in-/ output parameters)
Challenges
14. Challenge 2-3:
Input/ output parameters:
which are the most relevant?
Individual characteristics External training load
Internal training load
Recovery Psychosocial stress
Training effects
Physiological model of load capacity
How to achieve optimum performance in top sports?
Complex and depends on many factors
15. Physiological Model
Achieved
Work in progress
Phase 1-3 until June 2019
• Investigated which parameters play a role in the detection of
overload (literature study).
• Per practice partner: measured & desired parameters report
with the results.
• Research on which parameters have added value for the
model, in particular research whether biomarkers add
predictive value.
Ongoing research:
• Which parameters are optimal for quantifying external load
in football players (distance, speed, etc.).
Phase 4 until January 2020
• Research into which parameters have added value for the
model (systematic literature research).
• Report with analysis of the different parameters. A selection of
different parameters for external, internal and subjective load.
16. Challenge 1 & 4:
1. Differences per club
4. Choice of ML algorithm
Achieved
Work in progress
Load Monitor
Phase 1-3 until June 2019
• A prototype application has been built that makes it possible to:
• Pre-processing on large quantities of input data
• Combine data from different teams
• Create multiple prediction models
• Deal with variation in input and output parameters
• Apply different machine learning algorithms
• Software for anonymization of data
• User interface
Ongoing:
• Experimenting with different ML algorithms
Phase 4 until January 2020
• Further research on coherent variables (data-driven)
• Analysis of prototype results (in cooperation with Groningen)
• Demonstrator
17. Challenge
5. Validation
Achieved
Work in progress
Validation
Phase 1 & 2 & 3 until July 2019
• Research into validation possibilities
• Experiment with software
Phase 4 until July 2020
• Validation
• Analysis of prototype results (Groningen & Saxion)
To be tested with standard fitness tests (interval shuttle run, the maximum walking test and VO2max
test) and the experienced degree of fatigue (Rating of Perceived Exertion (RPE) Scale: a scale of 6 - 20
(6 - no effort and 20 - maximum effort). This can be compared with predicted values from the model).
19. MoSeS
health
component
Health problems by firefighters
• Heat stress (high core temperature)
• Exhaustion
Core temperature
• Direct measurement is not applicable
• Applicable methods provide estimation based on:
• Ear temperature
• Skin temperature
• Heart rate (model of Büller (2013) using Kalman filter
• Validated in studies with first responders (Buller, 2015)
• T > 38,00C / 38,50C
Exhaustion
• Based on (individual) HRR (Heart Rate Reserve)
• Short/ medium/ long term effort
20. MoSeS
health
component
Realized
• Health Status Component
• Demonstrator app
• Integration Moses
Advantages - modular structure
• Sensors
• Indicators
• Decision making
Results published
E-society 2019 http://www.iadisportal.org/digital-library/moses-health-
monitoring-system-for-firefighters#
Saxion Postit https://postit.saxion.nl/resolver/getfile/4D598F0E-3067-4C15-
B8F80CAB9BB3EA08
Het doel is om de belastingsmonitor voor voetbal te ontwikkelen waar verschillende externe belastings-parameters (o.a. playerload, aantal acties per minuut, high intensity distance, afgelegde afstand), interne belastingsparameters (gemiddelde hartslag, hartslag exertion index, tijdsduur in hoge hartslag zones) en subjectieve parameters (Rate of Perceived Exertion en evt. Total Quality of Recovery en aanvullende vragenlijsten) input zijn voor het achterliggende computermodel.
Op basis van individuele data uit het verleden en de huidige trainingsbelasting een verwachte huidige interne belasting worden berekend, welke kan worden vergeleken met de daadwerkelijk gemeten interne belasting.
T1.1 Het testen van verschillende soorten modellen en algoritmen om de data te analyseren (zoals lineaire regressie, Bayesian regressie, support vector regressie). Verschillende modellen en verschillende combinaties van parameters zullen getest worden en tegen elkaar afgezet.
T1.2 Onderzoek naar welke parameters optimaal zijn voor het kwantificeren van externe belasting bij voetballers (afstand, snelheid, playerload, etc., zie Bijlage B laatste figuur). Onderzoek naar bepalen welke parameters van toegevoegde waarde zijn voor het model, in het bijzonder onderzoek of biomarkers (stof die wordt gebruikt als een indicator van een bepaalde biologische toestand) als voorspellende waarde toevoegen. We denken hierbij met name aan het meten van lactaat, creatinine, fosfaat, magnesium, natrium, kalium, chloride en sulfaat.