Change Matters: Medication Change Prediction with Recurrent Residual Networks

IJCAI 2021

Chaoqi Yang¹, Cao Xiao², Lucas Glass², Jimeng Sun¹

¹UIUC, ²IQVIA

1

Outline

• Background
• Problem Formulation
• MICRON Model
• Experiment
• Conclusion

2

Task: Medication Change Recommendation

3

diagnosis

procedure

medications

Visit t-1

diagnosis

procedure

medications

Visit t

+ / -

• New problem, opposed to full medication set recommendation

MIMIC-III inpatient setting

IQVIA outpatient setting

Outline

• Background
• Problem Formulation
• MICRON Model
• Experiment
• Conclusion

4

Formulation

5

Outline

• Background
• Problem Formulation
• MICRON Model
• Experiment
• Conclusion

6

MICRON Model

7

Patient Representation

8

Embedding table

Embedding vector

Patient hidden representation

Residual reconstruction

9

Prescription network

Two consecutive prescriptions

Health updates to medication updates

Medication reconstruction

sigmoid

Training Stage with Other Objectives

10

Drug-drug interaction loss

Binary cross-entropy loss

Multi-label margin loss

Inference Stage

11

STEP 1: Medication vector update

STEP 2: Addition and removal

STEP 3: Medication set update

Outline

• Background
• Problem Formulation
• MICRON Model
• Experiment
• Conclusion

12

Data set and Baselines

• Two datasets
• MIMIC-III inpatient data
• IQVIA outpatient data

​

​

​

​

• Baselines
• SimNN, DualNN, LEAP, RETAIN, GAMENet

​

​

​

13

Performance Comparison

• We observe that our MICRON model shows low DDI rate, better Jaccard and F1 score.
• MICRON also performs well in predicting addition and removal med set with efficiency.

​

​

​

14

Ablation Study on Model Components

• We find our MICRON model has different sensitivity towards each component.
• With all components, the MICRON model works the best.

​

​

15

Outline

• Background
• Problem Formulation
• MICRON Model
• Experiment
• Conclusion

16

Conclusion

• This paper identifies the medication change prediction problem, which is easier, compared to traditional medication prediction.
• We learn effective representation of changing health conditions and use a residual health representation module to sequentially update medications.
• We conduct extensive experiments on inpatient MIMIC-III data and outpatient IQVIA data and show that our model can outperform the recent baselines. We also demonstrate the usefulness of all model components.

17

Change Matters: Medication Change Prediction with Recurrent Residual Networks

Thanks for Listening!

Chaoqi Yang¹, Cao Xiao², Lucas Glass², Jimeng Sun¹

¹UIUC, ²IQVIA

18