SLIDE 1
emergency departments (EDs) provide one of the greatest bottlenecks in the hospitalisation process – (more acute since pandemic!)
Clinical Authors: David Eyre, Peter Watkinson Reactive Environment: - - PowerPoint PPT Presentation
Clinical Authors: David Eyre, Peter Watkinson Reactive Environment: - - PowerPoint PPT Presentation
Authors: Rasheed el-Bouri, Tingting Zhu, David Clifton Clinical Authors: David Eyre, Peter Watkinson Reactive Environment: emergency departments (EDs) provide one of the greatest bottlenecks in the hospitalisation process (more acute
SLIDE 2
SLIDE 3
Increased risk of adverse outcomes for patients.
Some countries impose financial repercussions on hospitals for long patient waiting times.
Patients remaining in the ED still need to be cared for. Acts as a closed-
loop and slows the entire process down even further.
SLIDE 4
SLIDE 5
Triage
SLIDE 6
Triage Investigation
SLIDE 7
Triage Investigation Assessment
SLIDE 8
Triage Investigation Assessment
Decision
SLIDE 9
Predict what type of ward a patient will be admitted to (seven classes).
The type of ward is used so that any ward in the hospital with that functional capability can be considered
An accurate answer as soon as the patient walks in is the most useful!
Medical Cardiac Neuro Trauma ICU Surgical O&G ED
SLIDE 10
Curriculum learning has improved the performance of many algorithms that are trained using gradient descent.
No real consensus on the best type of curriculum for a given problem
Can we tailor-make a curriculum? Not just for a task but for a model too!
SLIDE 11
Training a neural network is Markovian
network new network a new network b Train with batch a Train with batch b
SLIDE 12
SLIDE 13
Some similarity metric, 𝐼, that allows us to sort our training data set from
most complex to least complex, such that
In this work we use cosine similarity as 𝐼 for images and the
Mahalanobis distance as 𝐼 for categorical and numerical data
𝐶𝑗 𝐶𝑘
𝐼 𝐶𝑗 < 𝐼[𝐶𝑘]
SLIDE 14
Weights 𝑋𝑗𝑘 of size 𝑁𝑗 × 𝑁𝑘 Reference vector of unique
elements, 𝑏
|𝑋𝑜
𝑗𝑘. 𝑏| and ∠ (𝑋𝑜 𝑗𝑘 . 𝑏)
𝑋𝑜
𝑗𝑘. 𝑏
SLIDE 15
Representation of layer 𝑗 is and all the layers are concatenated together. The full representation of a student with k hidden layers is
SLIDE 16
r Student state 2 1
DDPG or DQN
r = ∇𝑢𝑠𝑏𝑗𝑜𝑗𝑜 ∗ ∇𝑤𝑏𝑚𝑗𝑒𝑏𝑢𝑗𝑝𝑜
The actor (teacher) has two
- utputs :
- 1. Curriculum index
- 2. Batchwidth
SLIDE 17
SLIDE 18
SLIDE 19
SLIDE 20
SLIDE 21
Teacher learns how to degrade
performance in order to start again and achieve a better performance after training again
Bottom plot: orange is first output
index of teacher (index in curriculum), blue is second (width)
SLIDE 22
Plot shows performance on MIMIC-III
mortality prediction task
The teacher uses the same strategy
as on the ward admission dataset and achieves a strong performance for the student
Some metric of task similarity will
allow teacher transfer for training
SLIDE 23
Discussion of curricula learned for all tasks Constrained teacher Policy calibration Policy Transfer between Tasks Convergence of teacher selection
SLIDE 24
Feel free to contact me at rasheed.el-bouri@eng.ox.ac.uk with any questions
- r ideas