1. Applying Deep Neural Network to Enhance EMPI Searching
Stanley Liang
Supervisor: Prof. Jimmy Huang
School of information technology

2. What is EMPI EMPI is the acronym of enterprise master patient index
EMPI is a database used in a healthcare organization to maintain the data consistency of a variety of EHR information on the patients seen and managed within its various departments.
An identical patient has is likely to have different information regarding the same domain in different information systems.
EMPI plays a role as a universal identifier or label of an identical patient is the complex EHR system.

3. Measure of Entry Errors
Incorrect search results are most likely to be caused by entry typos
Edit distance, or Levenshtein distance, is a way to measure the dissimilarity of two words by counting the minimum number of changes to make them the same
Edit distances is commonly applied to natural language processing for automatic spelling correction.
To retrieve the due EHR information, we apply the similar idea to EMPI searching to classify the typos into three categories:
Match – no error
Minor error – 1 ~ 3 errors
Major error – 4 or more errors
A deep neural network model is applied in this task.
The algorithm can be implemented either by MATLAB or by C#.

4. Deep Learning
Deep learning generally means a neural network having more than two hidden layers for performing either supervised or unsupervised learning
Deep learning is commonly applied to pattern recognition in unstructured data, e.g. text, photos, videos, and sound
Deep learning is capable of discovering intricate structure in large data sets by the back-propagation algorithm
Deep learning provides extremely accurate predictions

5. The Feedforward-Backpropagation learning Cycle

6. Experiment Environments
MATLAB
MATLAB 2016b
Neural Network Toolbox
20,000 EMPI Records with random errors
C# & Visual Studio
Accord.NET Framework

7. Implementing Deep Learning in EMPI Information Retrieval
The manual input queries by the users usually have edit errors.
Error in text editing: Stanley  Stanly
Logical error: Female Male
Date error: 
Problem: if we use SQL query only, the machine cannot learn because SQL only use equal search and range search. If the patient is in the system but the query is wrong, it is difficult to retrieve the information or at high cost of computing time
Task
Provide adaptive search augmented by deep learning implicit search with good tolerance of occasional manual entry errors
Improve system efficiency

8. Algorithm Implementation
Levenshtein Distance
Deep Learning

9. Data set for testing Category Number of records Label description
No error
6,000
[1 0 0]
Totally correct
Minor error
1,000
[0 1 0]
ED=1 in a single column randomly, Total=1
ED=2 in a single column randomly, Total=2
ED=3 in a single column randomly, Total=3
ED=1 in two single columns randomly, Total=2
ED=1 in three single columns randomly, Total=2
ED=1 in a random column and ED=2 in another column randomly, Total=3
Major error
2,000
[0 0 1]
ED>=4 in a random column, Total>=4
ED>=4 in total in two random columns, Total>=4
ED>=4 in total in three random columns, Total>=4

10. Ten-fold cross-validation

11. Confusion Matrix Green Cells represent correct classification
The accuracy is 97.4% in this case
Class 1: match
Class 2: minor error
Class 3: major error

12. Conclusion
The deep neural network classifier has stable performance on this work with the average accuracy at 96.3%
We conclude that deep learning provides a good artificial intelligence solution for the EMPI query classification to enhance the overall system performance
We expect that artificial intelligence (AI) powered by deep learning will play an important role in a variety of healthcare decision making process

13. The End of Presentation
Thank you