TWIML #316 – Bridging the Patient-Physician Gap with ML and Expert Systems

Episode #316 of the TWIML AI Podcast is with Xavier Amatriain, Co-founder and CTO of Curai. They discussed the use of Curai’s machine learning models to help improve and augment primary care ultimately resulting in less workload for doctors. Xavier claims their models are as good or better than the “average doctor”.

Curai use expert systems for medical data and knowledge combined machine learning techniques in order to make “better” or more flexible decisions than the typical if / then rules that expert systems provide. Expert systems do not handle “noisy data” well but they can be trained with ML techniques to handle the “noise” of incomplete or incorrect data.

Their aim is to provide chatbot technology that could interact directly with patients in place of a human doctor or as tool to assist a human doctor while they talk with a patient in real time over the phone. Their technology would aim decision making by parsing the latest research, EHRs and test results to provide suggested potential illness or procedures based on the patient’s symptoms. The chatbot only route could pass the patient onto a live doctor for serious issues or refer them to another source of health care for minor complaints.

Listening to the podcast a couple of things sprung to mind about the general realm of medical chatbots and things like “autocomplete for doctors”.


Ah the notorious world of medical billing! An iPhone will correct your spelling of a brand (a “named entity”) in an iMessage (Android likely does the same). What’s to prevent something similar happening with “autocomplete for doctors”? The software could weight billable options higher over other cheaper or non-billable options.

Hack the chatbot

Once systems like these are live, how long will it be until someone figures out how to hack a medical care chatbot? With the right wording of their symptoms could someone get medical cert or sick note to get off work without needing to see a human?

Companies can obviously require examination by a human doctor but, for example, with a lot of clinics in Ireland not taking on new patients because they are already overloaded and understaffed, why wouldn’t clinics off load the work to a bot? And I know the obvious answers is “don’t build bots that have that functionality” but you know someone will attempt it for a competitive advantage. A clinic might even pay for the bespoke functionality.

Get In Touch