Informed Consent for Algorithmic Care: What Patients Aren't Being Told
Informed Consent for Algorithmic Care: What Patients Aren't Being Told
Informed Consent for Algorithmic Care: What Patients Aren't Being Told
A patient arrives at an emergency department with a fever and elevated heart rate. Within minutes of triage, an algorithm has scored their sepsis risk. That score determines how quickly they see a physician, which investigations are ordered, and whether they are admitted. The patient does not know this. Nobody tells them.
This is standard practice in hundreds of hospitals across North America. Epic's sepsis prediction model, various proprietary early warning scores, and triage algorithms all run silently in the background. They shape care. Patients consent to treatment, to procedures, to surgeries. They do not consent to algorithmic decision-making because they are never told it is happening.
The Consent Gap
Informed consent has four requirements: disclosure, comprehension, voluntariness, and capacity. A patient must be told what is being done, must understand it, must agree freely, and must have the cognitive ability to do so.
Algorithmic care fails at the first requirement. We do not disclose it.
Consider medication dosing. Pharmacogenomic algorithms adjust drug doses based on genetic profiles. Clinical decision support tools suggest insulin regimens and anticoagulation targets. These are recommendations a physician can override, but in practice the default is powerful. Most physicians follow the algorithm's suggestion most of the time. The patient sees a prescription. They do not see the computation behind it.
Or consider triage. When Babylon Health's chatbot directed patients toward self-care or emergency evaluation, the patients may have understood they were talking to software. But did they understand how it made its determination? Did they know what data it used, what conditions it was trained to detect, what its known failure modes were? We require that level of disclosure for surgical procedures and clinical trials. We do not require it for algorithms that make equivalent decisions.
Why This Is Hard
I recognize the practical objection immediately. A physician cannot explain every algorithm running in a modern EHR. There are dozens. Some are simple (BMI calculations, drug interaction checks). Some are complex (sepsis prediction, readmission risk scores). Requiring individual consent for each one would paralyze clinical workflows.
That objection is valid. It is also insufficient. The difficulty of doing something does not eliminate the ethical obligation to do it.
What Real Algorithmic Consent Could Look Like
I think the solution is tiered disclosure.
At the system level, hospitals should publish a registry of all clinical algorithms in active use, publicly accessible, including each algorithm's purpose, data inputs, accuracy, and known failure modes.
At the encounter level, when an algorithm materially influences a clinical decision, that should be communicated during the encounter. "Your sepsis risk score is low, which is one reason we feel comfortable managing this in the community." That takes ten seconds. It respects the patient's right to know what is shaping their care.
At the consent level, when an algorithm is the primary decision-maker (as with IDx-DR), the consent process should mirror procedural consent. The patient should be told what the system does, how accurate it is, what happens if it is wrong, and what their alternatives are.
This matters because autonomy requires information. A patient cannot exercise meaningful choice if they do not know how decisions are being made. The Canterbury v. Spence standard in US law established that patients are entitled to information a reasonable person would consider material. I find it difficult to argue that "a computer algorithm determined your treatment plan" is immaterial.
We built consent frameworks for surgery, for research, for data collection. We have not built one for algorithmic care. The algorithms are already here. The consent process needs to catch up.