Belitsoft was approached by the owner of a private medical center chain from the USA. Doctors and nurses working in his company were spending too much time on EHR-related tasks. This meant they either had less time for patients or more overtime work. Implementing speech recognition could be a solution to the problem - by talking to the machine the medical professionals could enter information quicker and even do it while examining the patient.
The client chose Belitsoft for the following reasons:
- Proven experience in working with both speech recognition technology and Healthcare domain;
- Competitive pricing;
- Good reviews and references.
The scope of the project was extensive and subject to change. So the client and us agreed on Agile development methodology along with the time and material cooperation model.
The client’s representatives visited Belitsoft office in Minsk before the kickoff to get acquainted with the team and the company leadership. Over the course of the project several key staff members went to the client’s office in the US to learn more about the end users and their work environment.
The development process was split into a number of 3-week sprints. Each sprint ended with a demo session where we showed the result of our work to the customer. These meetings were useful as a source of feedback for us and as a proof of money well-spent to the client.
The client has also put together a focus group from medical professionals working for him. These people proved invaluable in understanding the end users’ needs and testing of the features.
Belitsoft has developed a speech recognition system integrated with our client’s EHR. It was built as an on-premise solution due to security concerns.
Its most notable features included:
- Voice input of text and numbers;
- Voice commands for navigation inside the system;
- Automatic expansion of medical acronyms and abbreviations;
- An option of adding more dictionaries for medical specializations;
- An option to adapt to the voice of a specific medical professional.
The first release included three core dictionaries:
- general medicine
Each contained the data the system needs to recognize and process the words relevant to the appropriate niche. The “general medicine” dictionary was useful for all the fields within medicine, while “pathology” and “CT/MRT” had relatively few words and were cost-effective to implement. The system also included the option to expand the dictionary list, as mentioned above.
As one of the customer’s requirements we have also created an open API for the system to make it easy to integrate with other medical solutions.
As a part of this project, we have also found the most suitable headset for doctors and nurses. It had to be convenient enough to be worn 8 hours a day and provide high signal quality.
Despite the unexpected issues that have appeared over the course of the development, the system has successfully solved the customer’s problems.
Time spent on clerical tasks has decreased by 23%. The results were even better with older doctors, who were experts in medicine, but not as well-versed in computers. Moreover, the focus group has reported higher satisfaction with their work environment.
- Technologies: C++, Machine Learning, Neural Networks, SVM, Regression