Machine Learning for Health Care

There has been a constant effort in health care to use more information technology to improve care and outcomes.  Artificial intelligence and machine learning are two of the recently heavily-promoted technologies and an article in the New England Journal of Medicine discusses their potential for medical uses.    (NEJM Article)  Both concepts have at their core the notion that computers are more capable than humans of taking vast amounts of data and teasing out new patterns.  These machine learning programs basically create, test and constantly refine a model as new data is entered.  The end point supposedly will be a situation where a computer basically tells a clinician what the best treatment for a patient is, or maybe we just eliminate the clinician.   Obviously there are some critical assumptions in using machine learning.  One is that all necessary data is available and that it is accurate.  Good luck with that in health care.  Another is that all health situations are not unique; that is that there are always patterns that are relevant to every patient; you just have to find them.  For very complex cases, I am not sure that is an accurate assumption either.  Some machine-learning use cases are more purely administrative, for example language translation, finding data in existing health records or turning recorded medical notes into written ones.  But others are heavily clinical, intended to aid in diagnosis and treatment selection.  On the one hand, clinicians make plenty of mistakes today relying on their own judgment and experience and existing information technology, so maybe computers won’t be any worse.  On the other hand, caution in relying on computers would certainly be warranted based on past experiences (EHRs, for example, have definitely not provided the advantages proponents claimed they would) and most people might prefer mistakes made by a human as opposed to a machine.  Another critical factor to consider is the cost of incorporating even more technology into health care processes.  That cost is not just for hardware and software but more importantly for the impact of changing clinical workflows and data gathering.  As we have seen with EHRs, the hassles associated with use may outweigh the benefits and lead to greater clinician job dissatisfaction.