Patient, the (artificial) doctor will see you now

In the midst of all the hype about ChatGPT, Google and Deepmind quietly released Med-PaLM (Medical Pattern Language Model), an open-source Large Language Model, designed to respond to medical-related questions asked by either clinicians or patients. The platform provides immediate access to large amounts of structured and unstructured complex healthcare data, using machine learning algorithms to analyse and identify patterns and relationships in data that might not be immediately apparent to healthcare providers previously.

The Med-PaLM platform has the potential to deliver benefits to healthcare providers and patients, enabling clinicians to gain better insights into patient's health, identify potential risk factors for various conditions and make more informed decisions about patient care and more accurate diagnoses with the potential to improve patient outcomes.

However, Med-PaLM comes with potential risks. One of the biggest risks is the risk of bias or inaccuracies in the underlying data and algorithms, which would lead to inaccurate or unfair predictions if the data set is biased or incomplete. This could result in negative consequences and even harm to certain individuals or populations. Another risk is data security, as the platform handles a significant amount of sensitive healthcare data, including personal health information, there is the risk of it being mishandled or misused, compromising patient privacy. Additionally, the platform is susceptible to technical issues such as software bugs, system failures, and cybersecurity breaches, which could compromise its accuracy and availability. Misinterpretation of data is also a risk, as machine learning algorithms can sometimes make mistakes, as can humans, leading to potentially harmful decisions if clinicians rely too heavily on the platform without fully understanding its limitations.

As with any new technology, it will require more research and testing to determine its effectiveness and improve its performance to the point that it can match that of clinicians. Early testing results published by the developers demonstrated incorrect retrieval of information in 16.9% of responses, incorrect reasoning in around 10%, and inappropriate or incorrect content of responses in 18.7%. These figures are significantly higher than those for human clinicians, who have less than 4% incorrect retrieval of information, 2% incorrect reasoning, and 1.4% inappropriate or incorrect content of responses. However, the team is working on techniques to improve its performance, such as instruction prompt tuning and using interaction examples to produce more helpful answers as well conducting rigorous testing and validation and addressing bias in the data. More research and testing will be required to determine its effectiveness in the long term, and it will likely take time to fully understand its strengths, limitations, and potential healthcare applications.

Right now, Med-PaLM has the potential to assist clinicians in making more informed decisions about patient care. However, it cannot replace the judgment, expertise, and experience of a human clinician or make clinical decisions on its own. Clinicians will still need to use their clinical judgment and experience to interpret the data provided by Med-PaLM and make the best decisions for their patients. It remains important for both healthcare professionals and patients to be aware of these risks and to use the platform with a degree of caution and scepticism.

Overall, the Med-PaLM platform is a significant step forward in the application of AI in healthcare with huge potential to improve the quality of care and patient outcomes through simpler access to vast amount of healthcare information. However more work is required to provide assurance of the safety and validity of the information it presents....in the meantime, proceed with caution.