Google is testing Med-PaLM 2, a chatbot trained on a series of demonstrations by medical experts to hold a conversation about healthcare-related issues.
The company began testing the system with patients at the Mayo Clinic in April: The algorithm can generate answers to medical questions and perform automated tasks, such as summarizing documents or organizing reams of health data obtained from various sources.
These types of algorithms form the basis of future preventive medicine systems that will characterize healthcare in the developed world: Thanks to wearables and simple everyday devices, patients are constantly generating vast amounts of data about their health parameters, which feed them into Algorithms capable of generating data. The equivalent of a digital twin, a picture of their health that can be used to assess the need for more intensive testing or examinations.
In principle, Med-PaLM 2 is a large language model (LLM) trained exclusively using health data to answer questions ranging from diagnoses to symptom queries. The decision to use the canonical data formulated in the medical qualifying exam questions was just a matter of caution: training it on real data from medical records raises many confidentiality and privacy issues. However, this regulatory problem can be easily overcome once sufficient anonymity systems are in place and regulators realize that processing real patient data at scale can lead to enormous advances in medical science.
We are at an extremely important moment in the history of medicine: from palliative medicine, which works only on the basis of visible symptoms, to true preventive medicine, in which patients don't need to realize they have a problem: it's just one leap forward. The algorithm needs to detect something abnormal and evaluate it as worthy of further, more detailed examination so that their doctor can call them, and the doctor can receive full information on what triggered the alert before seeing them and request Conduct relevant tests.
In fact, wearables should never be interpreted by the user, who often lacks the proper knowledge and experience. These devices and other diagnostic tests that are performed on a daily basis are actually designed to support algorithmic systems that can maintain active monitoring and initial diagnosis in anomaly detection mode. Systems based on generative algorithms provide feedback and care, but the real question is being able to reduce patient suffering and healthcare costs, unlike other types of algorithms that focus on prediction