AI decision support in NICU

Question 1

Out of 140 million births every year, how many will require special newborn care?

Answer 1

30 million

Question 2

Out of 140 million births every year, how man will require intensive care?

Answer 2

8-10 million

Question 3

Reasons why a baby might end up in NICU

Answer 3

Difficult delivery
Birth asphyxia
Infection
Prematurity

Question 4

Common cause of injury at birth?

Answer 4

Hypoxic Ischemic Encephalopathy (HIE)

Question 5

How does HIE cause damage to the newborns brain?

Answer 5

Via O2 deprivation, compounded by a decrease in blood flow to the brain

Question 6

25% of newborns who suffer from HIE will develop severe neurodevelopmental disorders like…

Answer 6

Epilepsy
Cerebral Palsy
Developmental delays
Cognitive impairment

Question 7

What % of newborns die from HIE?

Answer 7

20%, it is the leading cause of death for newborn infants

Question 8

Chance of brain injury for preterm infants <32 weeks

Answer 8

1:3

Question 9

Chance of brain injury for preterm infants <28 weeks

Answer 9

1:2

Question 10

How common are preterm births

Answer 10

2.3 million each year born <32 weeks

Question 11

Options for neuromonitoring

Answer 11

EEG - measures cortical function
MRI - structural information
NIRS - Cerebral oxygenation
Cranial ultrasound - to detect IVH

Question 12

Why EEG is suitable for NICU?

Answer 12

Detects infant seizures (from stroke, asphyxia, etc)
→ Can cause further brain damage and require prompt treatment but often have no clinical signs
Can be used to inform diagnosis- HIE, IVH, stroke, etc.
Can predict future conditions
→ Can predict long-term neurodevelopmental outcome
→ Useful for early intervention programmes

Question 13

Limitations of EEG

Answer 13

Complex waveforms which change over time, and change between babies
Requires specialist expertise E.g. clinical neurophysiologist
Limited expertise available
Not 24/7
Monitoring not in real time

Question 14

Solution to EEG limitations

Answer 14

using AI to interpret it and raise the alarm if risk increases past a certain threshold
E.g. seizure detection
Can be implemented cotside in real time
Integrated into existing technology (i.e. EEG machine)
Scalable
24/7

Question 15

Quantitative EEG (qEEG)

Answer 15

Summarises the information produced by EEG as numbers
→ To do this we need to know what information it is important to summarise
This information is categorised into clear “features” in order to form the algorithim

Question 16

qEEG features

Answer 16

• Time-domain features:
  → Power (amplitude)
  → Statistical measures: e.g. standard deviation, skewness,
  kurtosis
  → Features from range-EEG (measures peak-peak amplitude)

- Frequency-domain features: 
→ Frequencies are filtered into four relevant frequency bands
     - Spectral power
     - Relative spectral power
     - Spectral shape

• Connectivity features:
  → Correlation, Coherence
  → Global or inter-hemispheric connectivity

Question 17

Pros of qEEG

Answer 17

• Does not require expert interpretation of the EEG
• Objective measures
• Easily scaled to analyse multiple EEGs
• Implemented on computer
• May uncover unknown/unseen characteristics of the EEG- features difficult to detect with the naked eye

Question 18

Cons of qEEG

Answer 18

• Sensitive to artefacts- infant motion, dirt, interference
• May be difficult to define a feature set to represent EEG
• May be difficult to interpret the results – difficult to relate some
  features to underlying physiology
• Lack of standards in how these measures are defined

Question 19

What is Artificial Intelligence?

Answer 19

Computer algorithms that ‘learn’ from patterns in data and then can draw inferences from new data, just like humans
→ Considering ‘supervised learning’

Question 20

Purpose of algorithms?

Answer 20

Classification (e.g. dog, cat, or horse)
→ Sorts categorical variables
Regression (e.g. house prices)
→ Predicts continuous variable

Question 21

Building a machine learning model

Answer 21

1. Prepare data
   - EEG: select clean (artefact-free) segments of the data
   (‘epochs’)
2. Feature set
   - Define which features to use
3. Select ML algorithm to use
   - Simple or more complex (e.g. linear versus non-linear)
4. Train
   - Defines the ML model
   - Include labels (e.g. brain injury, yes/no)
5. Test
   - With data different to training data
   - No labels- we’re trying to get the machine to add these itself
   - Assess performance with whatever metrics. Is it good
   enough?
6. Repeat/refine
   - Trial and error

Question 22

Data for training and testing a ML algorithm

Answer 22

Three different approaches to dividing up initial dataset into training and testing

1. 80:20 split
   80% for training; 20% for testing
2. K-fold cross-validation
   Divide data into K groups
   K=5: 4/5 groups used for training; ⅕ groups used for testing
   Repeat 5 times, once for each group
   K=10: 9/10 for training; 1/10 for testing; repeat 10 times
3. Leave-one-out cross-validation
   K-fold cross-validation to its logical extreme: using each data point in the dataset as a group
   E.g. EEG from 50 babies; train with 49 EEG; test with 1; repeat 50 times

Question 23

Overfitting in AI

Answer 23

Method performs well on train data but not on test data
• Poor generalisation
Common mistake!

Question 24

Solutions to overfitting

Answer 24

Need proper training AND testing

Change ML model (e.g. to a simpler model ie linear)

Question 25

Underfitting in AI

Answer 25

Performance will be roughly the same for train and test data

Poor generalisation also as underperforms

Question 26

Solutions to underfitting

Answer 26

Change ML model (ie to a more complex model)

More data?

Question 27

Examples of AI models

Answer 27

Linear & Logistic regression
Decision Trees
Neural Networks
Gradient boosting machines
Support vector machine