When the Machine Disagrees: Patient Autonomy vs. Algorithmic Recommendations
When the Machine Disagrees: Patient Autonomy vs. Algorithmic Recommendations
When the Machine Disagrees: Patient Autonomy vs. Algorithmic Recommendations
Here is a clinical scenario I have encountered in various forms. A 74-year-old man with metastatic colon cancer has decided, after long conversations with his family and oncologist, that he wants to stop chemotherapy and transition to palliative care. He is tired. He has thought about this carefully. His decision is informed and voluntary.
The hospital's clinical decision support system disagrees. Based on his performance status, tumor markers, and response to previous cycles, the algorithm recommends continuing to a third-line agent. The recommendation appears in the chart. It generates a flag.
Now the oncologist is caught between two authorities: the patient's expressed wish and the system's calculated recommendation. In theory, this is simple. Patient autonomy wins. In practice, it is more complicated.
The Weight of the Default
Clinical decision support tools are designed to influence behavior. They reduce variation, increase adherence to guidelines, and catch errors. But the same design features that make them effective at standardizing care make them coercive when a patient's preference diverges from the recommendation.
When a system flags a deviation, it creates a documentation burden. The physician must explain why they did not follow the suggestion. In malpractice terms, this is a record that the physician was warned and chose differently. If the patient's outcome is poor, that flag becomes evidence. The algorithm's recommendation, even when both physician and patient rejected it, has a legal afterlife.
I have spoken with colleagues who admit that the overhead of overriding a recommendation sometimes tips them toward following the algorithm, even when their clinical judgment says otherwise. The system is functioning as designed. It is also eroding physician and patient autonomy at the same time.
A Second Scenario
Consider a 55-year-old woman with newly diagnosed atrial fibrillation and a CHA2DS2-VASc score of 2. The algorithm recommends anticoagulation. She has read about the bleeding risks. Her parent had a devastating intracranial hemorrhage on warfarin. She wants rate control alone with monitoring. Her absolute stroke risk is approximately 2.2% per year. She understands this.
Is her choice irrational? The evidence supports anticoagulation at a population level. But she is not a population. She is a person with a specific fear grounded in a specific experience, making a risk calculation that weighs her values differently than the model does.
A good physician respects this. A system built around algorithmic adherence metrics does not.
The Physician in the Middle
When a physician recommends chemotherapy and the patient declines, two humans can negotiate. The physician can say, "I understand. Let's revisit this in two weeks."
An algorithm does not negotiate. It generates a recommendation based on its inputs. It does not hear that the patient is exhausted or see the family in the room. The recommendation persists, unchanged, regardless of the conversation at the bedside.
Where I Think the Line Should Be
Algorithms should inform, not adjudicate.
Override documentation should be simplified. A physician who respects a patient's informed refusal should not face a greater documentation burden than one who follows the algorithm. The current asymmetry creates a structural bias toward algorithmic compliance.
Patient preference should be a recognized input. If a system can incorporate lab values and imaging results, it can incorporate a documented patient preference. A system that recommends chemotherapy to a patient who has elected palliative care is failing to account for the most important variable in the decision.
Quality metrics should distinguish between unwarranted variation and patient-centered deviation. A hospital measuring physician "adherence" to algorithmic recommendations without accounting for patient choice is measuring the wrong thing.
The machine has an opinion. The patient has a right. When they conflict, the institutional incentives should make the answer obvious. Right now, they do not.