ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning...

17
ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials Novo Nordisk A/S: Sanna Herrgard Carla Gil Ingrid Holst John Holmer Nielsen Mattis Flyvholm Ranthe Ajser Serif Jesper Kjær Uppsala Monitoring Centre: Damon Fahimi

Transcript of ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning...

Page 1: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials

Novo Nordisk A/S:Sanna Herrgard

Carla GilIngrid Holst

John Holmer Nielsen Mattis Flyvholm Ranthe

Ajser Serif Jesper Kjær

Uppsala Monitoring Centre:Damon Fahimi

Page 2: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Dru

g D

isco

very •Drug candidate

identification

•Protein engineering

Dru

g Te

stin

g •Trial recruitment•Trial design and optimization -Adaptive Clinical Trials

•Clinical Processes -Medical Coding D

rug

Repu

rpos

ing •Finding new

therapeutic indications to already available drugs and chemical compounds

Machine Learning in Clinical Trials

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 3: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

• Increase automation level in the medical coding process

Assessment objective & scope

Objective

• Coding of concomitant medicationsScope

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 4: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Example output from coding

Verbatim Term Indication Drug Name (coded)

Drug Code (coded)

ATC code (selected)

Aspirin 81 mg tablet

to help prevent heart attack and/or stroke.

ASPIRIN 00002701004 B01AC, Platelet aggregation inhibitors excl. heparin

Aspirin Headache ASPIRIN 00002701004 N02BA, Salicylic acid and derivatives

Written by Investigator onConcomitant Medication Form

Outcome from coding

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 5: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Current concomitant medication coding process at Novo Nordisk

70-80%

20-30%

Single ATCcode

Multiple ATC codes

Auto-coding + manual coding

TMS

(62%) (38%)

Manual selection ofsingle ATC codeCRF form

Concomitant medication

Challenge: How to automate the entire process?

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 6: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Assessment of Koda – an automated coding engine from Uppsala Monitoring Centre

• Coding rules+ML• Training data: VigiBase

• ML: NLP + logistic regression• Training data: VigiBase

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 7: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

• Test Koda with NN test data sets

• Train Koda with NN training data sets

• Test Koda with NN training data sets àbaseline

• Performance of Koda prior to training

Export data 1

Train

3Test

4

Assessment of Koda – project plan

Baseline

2

• Export training and test data from Novo Nordisk (NN) clinical database

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 8: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Data sets used for training and testing of Koda

• Data from 12 different trials• Different therapeutic areas

• 70% of data used for training• 30% of data used for testing

Dat

a us

ed fo

r tes

ting

only

*

*No ATC codes available from NN

Num

ber

of c

onco

mita

nt m

edic

atio

nsML13: Assessment of Machine Learning Methods for Coding of Concomitant

Medications in Clinical Trials 11-Mar-2020

Page 9: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Results from Koda after training with Novo Nordisk data

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 10: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Drug Code coding efficiency

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 11: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Conformance of Drug Code coding with NN coding – high certainty predictions

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 12: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Analysis of disagreeing coding (sample of 181 conmeds with disagreeing coding)

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 13: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

ATC coding efficiency of Koda

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 14: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Conformance of ATC code coding – high certainty predictions

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 15: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

Summary from Koda assessment• ML supports automation of

concomitant medication coding

• Koda would enable NN to increase auto-coding rate from 62% to 79%• Koda provides single or multiple

suggestions for additional 15%

• Koda would enable NN to automate ATC selection

• High conformance with NN coding

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 16: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

• Understanding your data is critical when applying ML in clinical trials• Looking into percentages not enough

• Understand where your method may go wrong and address issues

Learnings

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020

Page 17: ML13: Assessment of Machine Learning Methods for Coding …ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020. Current

ML13: Assessment of Machine Learning Methods for Coding of Concomitant Medications in Clinical Trials 11-Mar-2020