A new Generative AI-driven service launched by AWS in 2023 is AWS HealthScribe. The service provides the ability to automatically generate clinical notes for documentation purposes based on patient conversations with the healthcare provider. Through a single API, conversations are seamlessly transmitted to the service as media files, triggering a process powered by Amazon Bedrock in the backend. This process automatically identifies speaker roles, classifies dialogues, extracts medical terms, and generates rich preliminary clinical transcripts and notes for the healthcare provider to approve. This service has been in the pipeline for a very long time and I am happy to see it finally come to fruition. I have been passionately interested in AI for medical applications, and many other companies have been working on something similar but with little traction. This service can have major impacts on delivering high-quality healthcare but also comes with challenges that need to be addressed before being used in any clinical setting.
Impacts in Healthcare
Reduced Patient Times
The substantial reduction in the time healthcare providers spend on writing and reading clinical notes could significantly streamline patient interactions. This is better for hospital systems that are struggling and want their healthcare providers to be more productive. HealthScribe would help physicians see more patients in the same amount of time because of the reduced burden of clinical notes. And with a little bit of rearchitecting this service would be a really good opportunity for a trailblazing entrepreneur to make AI scribes available to all healthcare providers.
Improved Patient Satisfaction
Patients would receive better healthcare primarily because of increased face-to-face interaction. HealthScribe also has the ability to provide a holistic picture of the patient's history because of its ability to process an abundant amount of patient history data and aid the caregiver in their decision with a potential diagnosis.
Reduced Rate of Errors
There would be a reduced rate of errors during the diagnosis because of a more thorough analysis of the patient history. Providing physicians with a more holistic picture of the patient proves to reduce the amount of errors during the diagnosis and improve patient outcomes. Notes from many years ago can be quickly analyzed and a more complete patient history can be provided within fractions of seconds compared to the arduous approach of reading through previous notes.
Challenges in Healthcare
Regulatory Challenges
There would be many potential regulatory challenges with HealthScribe completing sentences for physicians. There could be incorrect predictions that would be caused by incorrect training data and poor data provided during the inference step. As always, there will always be human error when approving the actions by the service and this would result in incorrect notes or incorrect diagnosis. Another unintended consequence would be if the same data is used from further training. This would further the error causing poor performance by the service.
Insurance Challenges
With auto-complete suggestions, there would be potential challenges with insurance pricing. Insurance will have to account for human and more importantly AI errors. An obvious solution would be to ensure that physicians approve every note entered into the system. This is already common practice and therefore would not change any existing process.
Hallucinations
Hallucinations are common with LLM models. It is completely possible that the service would hallucinate and provide incorrect predictions to the prompts provided. These would add to the workload for physicians if they have to correct the predictions of the AI and also complete notes for each of their patents.
In practice, there will be a fine line between completing sentences for physicians and leading physicians to diagnosis of the patient's condition. This will be a delicate balance that would need to be properly controlled by the service. If HealthScribe pretends to play doctor in real life it could spell trouble for the patient as well as the physician. An example of this occurs when HealthScribe completes sentences for the doctor, listing symptoms like fever, cough, fatigue, body aches, and respiratory issues for a patient. HealthScribe may incorrectly predict that the patient has COVID-19 when, in fact, it could also be the flu.
While it may not be immediately apparent, the landscape of AI in healthcare has become increasingly crowded, with no clear dominant player emerging. Navigating the complex terrain of healthcare politics poses a considerable challenge, particularly with hospitals already strapped for cash. Implementing change management is even more daunting. Ironically, AWS isn't the largest player in this space. It will be intriguing to observe how AWS HealthScribe will compete with larger players in the industry.
