Low adoption of decision support systems in medicine

Decision support systems are prevalent in almost every aspect of decision making. From loan repayment calculators to ovulation apps for pregnancy and from architectural planning to driving cars, decision support systems have become common place and the machine learning has percolated through them. In medicine, decision support in general, and therefore machine learning too, has been less successful but not for the lack of decision support apps, but for the lack of adoption. Even when decision support systems are in force, override levels are very high so what is going on?

The variety of decision support systems that exist means that different answers are required to explain this. It is also important to note that, despite the high rate of failure, some decision support systems are successfully used to the benefit of patients and clinicians. We can learn from successful systems by noting some patterns such as:

• Alerting systems that provide too many false positives are of no use. Like the Boy Who Cried Wolf, such systems are quickly ignored (or switched off).
• Alerting systems that tell you what you already know are just a nuisance.
• Alerting systems that don't alert you when they should, do more harm than good.
• Risk calculators that provide probabilities are meaningless and/or distracting to almost all patients and most doctors don't know how to interpret them appropriately even when they think they can.
• Decision support designed to be used during a consultation (such as treatment selection) can interrupt the consultation. It is too late to educate either patient or doctor (there is no time for that) and it can disengage the doctor from the patient while they are reading off a screen.
• People are typically only prepared to trust systems they can understand and verify, especially around complex decisions with serious consequences. Machine Learning systems are rarely that.

The list goes on but I'll keep it short. Computational decision support systems in medicine have been around for over half a century. A tiny fraction of them have ever had meaningful use and a lot of the usability mistakes are still prevalent.

Further reading: Machine Learning, Predictive Analytics, and Clinical Practice: Can the Past Inform the Present?