Dipole : Diagnosis Prediction in Healthcare via Attention- based - PowerPoint PPT Presentation

Dipole : Diagnosis Prediction in Healthcare via Attention- based Bidirectional Recurrent Neural Networks Author: Fenglong Ma, Radha Chitta, Jing Zhou, Quanzeng You, Tong Sun, Jing Gao Source: KDD '17 Advisor: Jia-Ling Koh Speaker: Shih-Han Lo Date: 2018/06/05 1

Outline  Introduction  Method  Experiment  Conclusion 2

Motivation 3

Goal Electronic Learn Health interpretable Records representations Improve the accuracy, provide better interpretation (Especially when the length of the visits is large) 4

Basic Notations  All unique medical codes from the EHR data:  Sequence of visits from the n-th patient:  Binary vector of medical codes:  Category representation: 5

Outline  Introduction  Method  Experiment  Conclusion 6

Proposed Model 7

Visit Embedding  Given a visit 𝒚 𝑗 , we can obtain its vector representation as follows: 8

Bidirectional Recurrent Neural Networks  Structure of unidirectional RNN: 9

Bidirectional Recurrent Neural Networks  Structure of bi directional RNN: 𝑔 𝑔 10

Attention Mechanism 1. Location-based Attention: 2. General Attention: 3. Concatenation-based Attention: 11

Attention Mechanism  Attention weight vector 𝜷 𝑢 :  Context vector 𝒅 𝑢 : 12

Diagnosis Prediction  Attentional hidden state (attentional vector):  The attentional vector is fed to the softmax layer to produce ( t+1 )-th visit information: 13

Objective Function  Cross entropy to calculate the loss for all the patients: Ground truth visit Predicted visit 14

Interpretation  The top k codes with the largest values are selected: 15

Interpretation 16

Outline  Introduction  Method  Experiment  Conclusion 17

Datasets 18

Results of Diagnosis Prediction 19

Results of Diagnosis Prediction 20

Assumption Validation 21

Outline  Introduction  Method  Experiment  Conclusion 22

Conclusion  By employing bi directional recurrent neural networks ( B RNN), Dipole can remember the hidden knowledge learned from the previous and future visits.  Attention Mechanisms allow us to interpret the prediction results reasonably. 23