Electronic Health Records

Case Introduction:

The client had years of past Electronic Health Records (EHR). Specific and multiple information points from these EHR documents were to be converted in a meaningful dataset which would be of immense use. Performing optical character recognition (OCR) and extracting text with structure would significantly improve their searchability.

Further to above, the client also wanted to investigate the use of voice based navigation for shortening the data access from Electronic Medical Records (EMR) application.


The key objectives of the project were –

  • Text extraction: Extract the text from documents in different formats (searchable PDFs, scanned PDFs, image files, etc.)
  • Structure extraction: Extract the major structural components from the document (ex. Prescription, medical history, etc.).
  • Search integration with EMR application: Integrate the search functionality with the EMR application using elastic search.
  • Voice-based navigation: Perform automatic speech recognition (ASR) to navigate the EMR application.
  • Performed OCR on the complex (skewed, rotated) historical (poor quality) scanned documents.
  • Document modeling to extract the structure as required by the client.
  • Optimized the architecture and pipeline to return instantaneous search results using elastic stack.
  • Tuned and trained the open-source ASR engines for the specific use case.
  • Applied two-stepped ASR processing to optimize data transfer load for voice navigation.
  • The EHR documents had significant variations in format and quality (rotated, skewed, bad contrast, etc.) We had to tune-up the pre-processing pipeline to ensure good quality output.
  • The initial processing pipeline with python was not giving a satisfactory performance. We were able to extract 4x performance using PySpark parallelization.
  • The privacy of the doctor-patient conversation was of paramount importance. We had to come up with a two-stepped speech recognition engine to ensure this.
Solution Impact:
  • Achieved OCR accuracy of ~91%.
  • Achieved ASR accuracy of greater than 90%.
  • The search solution potentially helps the end-users (doctors) to eliminate 15-25% of manual effort.
  • The voice-based navigation would potentially help in improving the quality and efficiency of the patient-doctor encounter.
Technology Stack

AI/ML Development




Please drop us a message.