3D Sensing Algorithms Towards Building an Intelligent Intensive Care Unit

Preview:

DESCRIPTION

2013 Summit on Clinical Research Informatics

Citation preview

March 21, 2013

3D Sensing Algorithms Towards Building an Intelligent Intensive Care Unit

Colin Lea PhD Student Computer Science Johns Hopkins University

Suchi Saria James Fackler

Greg Hager Russell Taylor

preliminary work Automatic analysis of ICU workflow •  Gather data passively •  Develop/evaluate techniques for task recognition •  Summarize staff activity

Hundreds of tasks •  Checkups •  Emptying tubes •  Documentation

Benefits •  Safety •  Workflow Optimization •  Resource Allocation

2  

Task  Comple-on  Concept  Sketch  

safety and compliance Many reoccurring events!

Compliance •  Central line infections •  Ventilator-Associated Events •  Turning patients

3  

setup

4  

example depth video

challenges in an icu Difficult Environment •  Clutter •  Varied layouts

Little written direction •  Which nurse is doing

what? •  Hierarchy of staff

Cognitive load •  Too many hard tasks •  Inefficient process

Movable Equipment

Infant

Monitors Staff

6  

task analysis approach

7  

Actions: 1) Checking vitals 2) Documenting 3) Observing 4) Performing

procedures 5) Changing Foley

Catheter 6) Ventilator use 7) Other

action sequences

8  

Classification is function of action sequence Problem: Changing Dimensionality!

Variable event duration, people count

Terminology: •  Action: semantic label •  Action sequence: composed of a set of

segmentations for each individual in an event

=  

feature extraction

9  

Motion Statistics

Histogram of Orientations Interaction Coefficient

Virtual Proximity Sensors

experimental setup

10  

Johns Hopkins Medical Institute Pediatric Intensive Care Unit 5.5 hours of data 122 Action Sequences Annotations by the attending Single Kinect

Pediatric  ICU  dataset  

analytics

results

12  

Evaluate with:

Support Vector Machines Kernel = RBF, C =100 Decision Forests Max Features = 5 Trees = 20

Actions Samples SVM Accuracy (%)

Decision Forest Accuracy (%)

Documenting 11 54.5 63.6 Observing 47 70.2 83.8 Checking vitals 28 64.3 74.2 Performing procedures 14 42.9 37.1 Changing Foley Catheter 5 80.0 100.0 Using Ventilator 9 40.0 33.3 Other 8 50.0 55.0 Overall Average 60.6 69.5

Pediatric  ICU  dataset  

analytics

Most common: Checking vitals Most time per task: Procedures Nurse time: 25 min/hour

analytics

Most common: Checking vitals Most time per task: Procedures Nurse time: 25 min/hour

Pediatric  ICU  dataset  

Early  PICU  dataset  

average action locations

future work

Target specific actions Compliance Scale up Increase feature fidelity

16  

questions

Colin Lea colincsl@gmail.com

Johns Hopkins University www.colinlea.com

Suchi Sar ia James Fackler Greg Hager Russell Taylor

Recommended