Back to Success Stories
Health & Wellness

EHR/Diagnosis Automation

The client provides an electronic health record (EHR) SaaS platform for orthopedic medical practices to help organize and store patient documents and streamline workflows. The system is designed to record patient data seamlessly, create custom treatment plans, and provide orders for surgeries, prescriptions, and referrals with a focus in orthopedics.


The client manually entered patient data, assessments, doctors’ orders, and notes. This was a time consuming process subject to human error. In order to create a more efficient process, the organization sought to optimize the process and eliminate the risk of error. The process was automated by allowing doctors to speak patient summary forms right into the system.
Text from doctor-patient conversations was analyzed, recorded, and organized into 8 categories: Strength, Range of Motion, Palpation, Reflexes, Special Test, Sensation, examHeader, and Inspection. For instance, in text with 20-30 sentences, 1 or 2 might be categorized as “Reflexes” and others as “Inspection.”
Confidentiality and lack of data were initial roadblocks, and the organization thought the project was undoable due to lack of doctor-patient conversation audio training data.
Many organizations face similar challenges:
1. Uncertainty about how to leverage data:
Organizations have patient records and follow up data but are unsure how to use it to their advantage.
2. Limited knowledge of artificial intelligence:
Organizations looking to automate business processes may understand they need to use AI but don’t have the expertise on their team equipped to identify which problems are best suited to AI solutions.
3. Dealing with private data:
Hospitals may be open to providing companies with private data while others are not, and conversations between doctors and patients are confidential.


  • Save time
  • Automate processes
  • Avoid manual error
  • How to leverage and organize data in an actionable way


Within two weeks, Fusemachines assigned a PhD specialized in speech processing and two machine learning engineers. The Fusemachines team solved the challenge by developing an automated speech recognition (ASR) and natural language understanding (NLU) models that converted audio to text. Ultimately the NLU model extracts the musculoskeletal examination findings from the text, and into the electronic health records (EHR).
First, the client did not have all the necessary data to train a successful model so Fusemachines developed a method to create synthetic transcripts. This involved combining clinical findings from the EHR database with random sentences outside the domain which increased the amount of data available to train the model.
Next, a Bidirectional Encoder Representations from Transformers (BERT) was trained to perform named entity recognition (NER) using the synthetic transcripts as training data. The BERT model was trained to classify sentences into categories
In addition to standard named entity recognition metrics, to communicate the capabilities of the model, Fusemachines loaded results from the test data set into a annotation tool (BRAT) server to visualize which sentences are extracted.
Fusemachines created an API that ingested an audio stream and extracted entities. These extracted entities are returned in JSON format and saved in the appropriate fields in the electronic health records (EHR).

How was our solution beneficial?

To drive efficiency and revenue, healthcare organizations strive to optimize doctor-patient interactions. Medical professionals’ time is expensive and they need to meet patient workload. Doctors are better able to interact face to face with their patients when they are not splitting their attention between talking to patients and recording their information making interactions more personal. This system allows doctors to do their jobs in meaningful and effective ways. The speech recognition tool can integrate with most EHR systems. It can be adapted to any medical practice. It is faster and cost effective over manual transcription
Unlike many forms of AI, most people are familiar working with Siri giving an ease of use and design and a frictionless user experience. The system can be implemented and used immediately and doctors did not have to learn the technology. It is a user friendly system which is easily adaptable. It saves hours, frustration, and paperwork and is much more efficient.
Users don’t have to use cloud services to run the system Fusemachines built. Because of this, it works for governments and other companies that need to host models on premise. The ASR/NLU models were custom built to fit unique orthopedic conversations.

Leveraging Data

1. Collect data
Organizations looking to leverage machine learning and AI should establish a data strategy to retain or acquire the proper amount and type of data to facilitate accelerated research outcomes. Fusemachines developed a strategy to attain the necessary data and data infrastructure for the project.
2. Assess data
Feature selection is critical for the development of accurate machine learning algorithms and requires properly labeled data. features in the data set.
3. Pay attention to nuances
It’s crucial to understand what the data means, how the algorithms function, and identify any outliers to make sure there are no issues with statistical analyses.
Alternative Applications
  • Any manual entry process; anywhere you want focus on interaction not data entry, diagnostic/mechanical setting/heavy, industry specific nomenclature
  • Reduce friction of interacting with technology is a great place for NLP
  • Hospitals and healthcare organizations can detect and categorize symptoms of COVID-19


The client had a vision they didn’t think was possible. Fusemachines engineers integrated with their team, advised and consulted them, helped them overcome roadblocks, and delivered a demonstrable product in less than 4 months. This technology made doctors’ lives easier and saved valuable time. Fusemachines accelerated their product roadmap and are working with them to come up with more areas to deliver AI into their product roadmap. The customer support is great. Many got rid of our dictation service, which was costing them around $120,000/year. A doctor claims saving at least $1,000 per month on transcription costs.
Among the 15+ created, the most successful model performed with 85% accuracy classifying sentences into the 8 categories used to fill the patient’s online summary form.
We use cookies to improve your experience & analyze usage. By continuing, you agree to our Privacy Policy.