The press release tries to spin it very hard, and don’t miss the adjectives to make it sound “thrilling”! I honestly ran out of synonyms, but still I am linking here.
What is missing? In addition to needed technical advances, AI must meet the same bar for clinical evidence that is expected from other clinical interventions. For a given AI tool to be used, evidence that it will perform in a safe and effective manner must be demonstrated using randomized controlled trials designed to test the tool against an established standard. For the tool to be accepted, trials need to be designed to show that it provides a benefit to the patient, the patient’s family, the health professionals using the tool, the health care system at large, or some combination of these potential beneficiaries. Unfortunately, there is currently limited high-quality evidence of this type available for many medical AI systems.
They have expanded the scope at the outset:
NEJM AI will be an interdisciplinary journal facilitating dialogue among stakeholders invested in using AI to transform medicine. NEJM AI will intentionally pair “pre-clinical” and clinical articles to deliver critical context to both clinicians and non-clinician researchers. The journal will bridge the fast-moving developments in AI, informatics, and technology in medicine with the application of these advancements to clinical practice. NEJM AI will cover the application of AI methodologies and data science to biomedical informatics, connected health, telemedicine, medical images and imaging, personalized medicine, policy and regulation, and the ethical and medicolegal implications of AI.
I still have to get access to their articles but in case you wish to follow them on Twitter, here’s the link to NEJM AI.
This is a broad recognition that the “future” will be scripted along “engineering-medicine fusion” and “inter-disciplinary”. Beyond the call for “standardised” data-sets, its broad applicability/generalisability should be mentioned. Does it improve the quality of life or just make it more accessible? What is the return on investment? How do you calculate the compute cycles? And many more.
I have enough failures here to understand the complexities.