department of surgery and cancer imperial college london 20 may 2014 norman fenton queen mary...
TRANSCRIPT
Department of Surgery and CancerImperial College London
20 May 2014
Norman Fenton
Queen Mary University of London and
Agena Ltd
Improved Medical Risk Assessment and Decision-making
with Bayesian Networks
Overview
• Why Bayes?• Why Bayesian networks?• Why NOT learn the models from data only?• Case study• Challenges and conclusions
1. WHY BAYES?
The Harvard ProblemOne in a thousand people has a prevalence for a
particular heart disease. A test to detect this disease has:• 100% sensitivity• 95% specificity If a randomly selected person tests positive what is the probability that the person actually has the disease?
Bayes Theorem
E(evidence)
We now get some evidence E.
H (hypothesis)
We have a hypothesis H with prior probability P(H)
We know P(E|H) but we want the posterior P(H|E)
P(H|E) = P(E|H)*P(H) P(E)
P(E|H)*P(H)P(E|H)*P(H) + P(E|not H)*P(not H)
=
1*1/1000
1*1/1000+ 5/100*999/1000P(H|E) = =
0.001
0.001 + 0.049950.0196
Waste of time showing this to most people!!!
Slide 6
Imagine 100,000people
Slide 7
Out of whom100 has thedisease
Slide 8
But about 5% of theremaining99900 peoplewithout thedisease testpositive.That is 4995 people
Slide 9
So 100 out of 5095 who testpositiveactually havethe disease
That’s justunder 2%
That’s very different fromthe 95% assumed by most medics
Total people100,000
1/1000
999/1000
Have the disease100
Don’t have the disease
99,900
So 100 out of 5,095who test positive actuallyhave the disease, i.e. under 2%
Test positive100
Test negative0
Test positive4,995
Test negative94,905
100%
0%
5%
95%
2. WHY BAYESIAN NETWORKS?
A Simple Bayesian Network
..but here is a typical
causal model
Calculations from first principles are
infeasible and incomprehensible
Actual model in medical negligence case
This model already reaches limit of comprehensibility for
manual calculations and event trees
• MRA• CA
• Ischaemic• Small aneurysm• Large aneurysm• CSP
Detected by Test9,900
Undetected by Test100
Detected by Test90
Undetected by Test10
Detected by Test0
Undetected by Test10,000
Die from burst/bleeding
Die from CSP
99%
1%
90%
10%
50%
0%
100%
2%
2%
CA Test PathwayCause of Palsy Test Result Outcome Deaths
2
0
5000
5002TOTAL
= 1.495%
1
14,952 out of 1,000,000 give risk
Stroke
Strokes
Don’t die
99
Stroke
Stroke
Die from burst/bleeding
Don’t die Stroke
Don’t die Stroke
1%
1%
1%
1%
1%
50%
98%
98%
99
2
981
1
1
0
10 0 0
5000
500050
1%Stroke
50
97999799
9950
Total people1,000,000
Large9,900
Small100
CSP10,000
Others (ischaemic)980,000
1%
1%
1%
98%
Aneurysm10,000
99%
Total people1,000,000
Large9,900
Small100
CSP10,000
Others (ischaemic)980,000
Detected by Test9,405
Undetected by Test495
Detected by Test50
Undetected by Test50
Detected by Test9,000
Undetected by Test1000
Die from burst/bleeding
Die from burst/bleeding
Die from CSP
Die from CSP
1%
1%
1%
98%
95%
5%
50%
50%
50%
90%
10%
2%
2%
20%
MRA Test PathwayCause of Palsy Test Result Outcome Deaths
10
1
1800
500
2311TOTAL
= 0.2311%
0
0
0
2311 out of 1,000,000 give risk
Aneurysm10,000
99%
Much better solution…use a Bayesian Network tool
Computation for Catheter Angiogram
Mean:9950
Mean:5002
Computation for MRA Scan
Mean:0
Mean:2311
The Calculator Analogy
No need for p-tests or classical confidence intervals
• Drug “Precision” weight loss: Everyone in trial lost between 4.5 and 5.5 pounds
• Drug “Oomph” weight loss: Everyone in trial lost between 10 and 30 pounds
• Which drug can we ‘accept’, i.e. reject null hypothesis of ‘no weight loss’?
• Classical stats provides nonsensical answers
No need for p-tests or classical confidence intervals
3. WHY NOT LEARN THE MODELS FROM DATA ONLY?
A typical data-driven study
Age Delay in arrival
Injurytype
Brain scanresult
Arterialpressure
Pupildilation
Outcome (death y/n)
17 25 A N L Y N
39 20 B N M Y N
23 65 A N L N Y
21 80 C Y H Y N
68 20 B Y M Y N
22 30 A N M N Y
… … … .. … …
Delay in arrival
Injurytype
Brain scanresult Arterial
pressure
Pupildilation
Age
Outcome
A typical data-driven study
Purely data driven machine learning algorithms will be inaccurate and produce counterintuitive results e.g. outcome more likely to be OK in the worst scenarios
Delay in arrival
Injurytype
Brain scanresult Arterial
pressure
Pupildilation
Age
Causal model with intervention
Dangerlevel
Outcome
TREATMENT
..crucial variables missing from the data
Determining drug effectiveness
Basic results for drug effectiveness
Drug AThe mean financial benefit is $4156
Drug BThe mean financial benefit is $2777
Ban drug B?
Model with latent variable (same data)
Note that most patients in the sample had minor
case of the condition
…and most patients were given drug A
Results with 'Patient condition' major
Drug B30% positive outcome.The mean financial benefit is $1000
Drug AOnly 10% positive outcome.The mean financial benefit is $400
OK, so we might need expert judgment when we have missing data, but with good experimental design and lots of good quality data we can surely remove dependency on experts ……
A machine learning fableA and B are two medical conditions very well known to doctors Bill and Ludmila. These conditions are pretty rare (both have an incidence of about one in 1,000 people). There is a third medical condition C (whose name is “FiroziliRalitNoNeOba”) that Bill has heard the name of, but knows nothing about. But Bill has heard that patients with either A or B usually also have C. Bill has a massive database of 600,000 people with the details of which conditions they have.
Bill’s dataPatient number A B C 1 No No No 2 No No No 3 Yes No Yes 4 No No No 5 No No No 6 No No No 7 Yes No Yes 8 No Yes Yes 9 No No No 10 No No No 11 No No No 12 No Yes Yes 13 No No No 14 No No No …. … … …. … … 600,000 No No No
Bill’s machine learning mate FredCan use this database to ‘discover’ the underlying causal model (Bayesian Network) relating A, B, and C. But Ludmila says she knows the correct model without data:
Fred warns against this
She also “knows” the probability tables
Fred’s learnt model
• Ludmilla disagrees with the last column of table C • Fred: “Not enough data for that”• Bill: “…why can’t we simply conclude that C must be true when
both A and B are?”
600 out of 600,000 have condition A
600 out of 600,000 have condition B
Every single person with condition A also has C and every single person with B also has C.
Ludmilla’s knowledge
• The name of Condition C - FiroziliRalitNoNeOba - is actually a Russian word.
• Its literal translation is:– ‘A person suffering from either Firoz or Ralit but
not both’. – ‘Firoz’ is the Russian word for condition A and
‘Ralit’ is the Russian word for condition B.”
Moral of the story
• Sometimes you have to trust experts to provide more informed quantitative judgement than you will get from data alone.
• Even really big datasets will be insufficient for some very small problems.
• Trusting the expert can save you a whole load of unnecessary data-collection and machine learning effort.
4. CASE STUDY
Trauma Care Case Study• QM RIM Group
– William Marsh– Barbaros Yet
• The Royal London Hospital– Mr Zane Perkins– Mr Nigel Tai– ACIT Data
• US Army Institute of Surgical Research– Lower Extremity Injury
DataYet, B., Perkins Z., Fenton, N.E., Tai, N., Marsh, W., "Not Just Data: A Method for Improving Prediction with Knowledge", Journal of Biomedical Informatics, 2014 Apr;48:28-37
BN v MESS Score
• Prediction: coagulopathy, death (c.f. GCS, TRISS)• Flexible inputs• Patient’s physiological state
– Causal modelling: informed by knowledge
How the BN Model Differs
Life Saving: Prediction of Physiological Disorders
Limb Saving: Prediction of Limb Viability
www.traumamodels.com
5. CHALLENGES AND CONCLUSIONS
Challenges• Apparent paradox on using experts• Expert systems have a bad reputation• Resistance to subjective priors• Building new large-scale BN models, especially
with minimal data• Interacting with large-scale BN models• Explaining the results of BN models
BAYES-KNOWLEDGE (Effective Bayesian Modelling with Knowledge Before Data)www.eecs.qmul.ac.uk/~norman/projects/B_Knowledge.html
Conclusions (1)
• Purely data driven approaches using Machine learning and statistics DO NOT WORK
• At best captures what did happen Vs what would have happened
• Need to move to data + knowledge approach• BNs provide the key
Conclusions (2): BN Benefits
• Data + knowledge• Models uncertainty and causality• Predictions and diagnosis• Avoid medical statistics fixation on p-values
and confidence intervals• Incorporate qualitative and quantitative
variables• Identify causal effects without RCTs• New generation expert systems
Blatant Plug for Book
CRC Press, ISBN: 9781439809105 , ISBN 10: 1439809100