Apoio deciso em medicina intensiva usando ECBD

1
Apoio à decisão em medicina intensiva usando ECBD

• Pedro Gago I P Leiria

2
Intensive care

• About 250 variables are needed to describe an ICU
  patient
• Humans are unable to cope with more than seven
  variables at a time

3
Objectives

• Assist ICU doctors by providing accurate and
  timely predictions for
• the final outcome
• organ dysfunction or failure
• Must overcome natural physician resistance

4
Intensive Care Medicine

• Condition is severe to the point where it is very
  difficult for doctors to assess the patients
  state
• Objective is to stabilize in order to allow
  transfer to other units
• Highly invasive and very costly

5
Intensive Care Medicine

• Data from bed-side monitors may contain useful
  information
• Presently such data is not stored

6
Practical Issues

• Some variables values must be collected manually
• Urine output
• Data Quality
• Errors caused by human intervention
• Sensor malfunctions

7
Scores in use

• SAPS indicative of the patients condition
  severity
• The worst values the first 24 hours of stay in
  the ICU are collected and the score is calculated

8
Scores in use (2)

• SOFA measures organ dysfuntion/failure (worst
  daily values)
• Cardiovascular, hepatic, central nervous system,
  respiratory, renal, coagulation
• Worst daily values
• Indicative of patients condition evolution

9
INTCare

• Decision Support System to assist ICU doctors
• Uses available data in order to predict outcome
  and organ dysfunction/failure
• Not intended to replace doctors

10
INTCare (2)

• Semi-autonomous updates its models as new data
  arrives
• Performance expected to improve with time
• Better results through the use of real time data

11
Architecture
12
Architecture (2)
13
EURICOS II

• Data from 42 UCI from 9 countries
• 10 months (1998 and 1999)

14
EURICOS II (2)

• Data available includes
• case mix (age, origin, etc)
• SAPS score
• daily SOFA scores
• intermediate outcomes
• final outcome

15
INTERMEDIATE OUTCOMES
16
Ensemble

• Training
• Each model is trained on different subsets of the
  dataset
• Each variable has a 70 chance of being selected
• Starts with equal weights

17
Ensemble

• Evolution
• Results from batches of records
• Weight adjustments according to individual model
  performance
• Worst performing models are deleted from the
  ensemble
• New models are trained using the most recent data
  and included in the ensemble

18
Ensemble

• Preliminary results (evolution doesnt include
  new models)
• Ensemble trained with all cases still outperforms
  ensemble trained with less cases followed by
  weight adjustment
• Both outperform best individual model

19
Future Work

• Greater volume of data deployment in other
  ICUs
• Reduce prediction window (next 6 hours instead of
  next day)
• Suggest course of action (must be delayed until
  physician resistance is lowered)

20
INTCare

• Thank you.
• Questions?
• pgago_at_estg.ipleiria.pt