Predicting warfarin dosage from clinical data: A supervised learning approach

Intelligent Database Systems Lab

Presenter : CHANG, SHIH-JIE

Authors : Ya-Han Hu, Fan Wu a, Chia-Lun Lo, Chun-Tien Tai b

2012.AIM.

Predicting warfarin dosage from clinical data: A supervised learning approach

Intelligent Database Systems Lab

OutlinesMotivationObjectivesMethodologyExperimentsConclusionsComments

Intelligent Database Systems Lab

Motivation

Physicians use computerized dosing nomograms of warfarin as reference .It merely consider age and INR values not enough for dose adjustment.

Intelligent Database Systems Lab

Objectives• Build a warfarin dosage prediction model utilizing a

number of supervised learning techniques to help dose adjustment.

Intelligent Database Systems Lab

Warfarin

Intelligent Database Systems Lab

Prediction model for warfarin dosing- Single classifiers (1) KNN

(2) SVRGiven a set of training instances

xi : input vector yi : actual output of xi

a regression function

ε-SVR can be formulated

regression hyperplane

Intelligent Database Systems Lab

Methodology - Single classifiers (3) M5 (model-tree-based regression algorithm)

use standard deviation reductionTree-building :

specific node

standard deviation of the class values of all instances in a child-node Nt,i,

Intelligent Database Systems Lab

Methodology - M5

error term

tree-pruning

Intelligent Database Systems Lab

Methodology – MLP

(4) MLP

Intelligent Database Systems Lab

Methodology – Classifier ensemble

Voting (weight)

Bagged Voting method

Decide the estimated output by combining the results of different classifiers.

Intelligent Database Systems Lab

Experiments – Data preparation Collected 587 clinical cases (INR value 1~3)

Drug-to-drug interaction (DDI)424 163

Use Bagging496

Intelligent Database Systems Lab

Experiments – Performance measures

Experiments – Evaluation results

Intelligent Database Systems Lab

Experiments – The average of evaluation results

Intelligent Database Systems Lab

Conclusions– The investigated models can not only facilitate

clinicians in dosage decision-making, but also help reduce patient risk from adverse drug events.

Intelligent Database Systems Lab

Comments• Advantages

– More accurate.• Applications

– Warfarin dosage prediction.