Control systems

In one line

A control system measures something, compares it to a target, and acts to close the gap — then does it again, forever. When the loop includes a patient and no human in the middle, it is closed-loop, and that is where medicine's most interesting devices and its hardest accountability questions both live.

Open versus closed loop

Open loop — the device acts and does not check. A syringe pump set to 5 mL/h delivers 5 mL/h regardless of what happens to the patient. Simple, predictable, and entirely dependent on a human deciding the number was right.

Closed loop — the device measures the result of its own action and adjusts. A CGM reads glucose, an algorithm decides, a pump delivers insulin, glucose changes, the CGM reads again. Nobody asked permission for the second dose.

That difference is not incremental. In an open loop, a human is in the decision. In a closed loop, the human wrote the policy and then left the room.

PID, briefly, because it's everywhere

The workhorse controller is PID, and its three terms are worth knowing because they map to recognisable clinical instincts:

• Proportional — react to how far off you are now. Glucose is high, give insulin.
• Integral — react to how long you've been off. It's been high all afternoon, give more.
• Derivative — react to how fast it's changing. It's climbing steeply, act before it peaks.

Tune them badly and you get behaviour any clinician recognises: overshoot (too aggressive — hypoglycaemia), oscillation (chasing your own tail), or sluggishness (safe, useless). The artificial pancreas literature is largely the story of tuning that trade-off when the cost of overshoot is a seizure.

Why physiology is a hard plant to control

Control theory calls the thing you're controlling the plant. Human physiology is a spectacularly awkward one:

• Long, variable delays. Subcutaneous insulin takes tens of minutes to act — the digital twin lab models exactly this depot. A controller acting on 30-minute-old information, whose actions take 40 minutes to land, is steering a ship, not a bicycle.
• The sensor is indirect. Interstitial glucose is not blood glucose; it lags and drifts. You are controlling a variable you cannot actually see — and the sensor has its own failure modes.
• The plant changes. Illness, exercise, stress, hormones, a large meal. The system you tuned for is not the system you have today.
• Asymmetric consequences. Slightly high for an hour: unwelcome. Slightly low at 3am: potentially fatal. A controller that treats error symmetrically is wrong about the world.

That asymmetry is why real medical controllers are deliberately conservative, and why "it performed better in simulation" is a warning rather than a result.

Where you meet them

Ventilators (pressure and volume regulation), infusion pumps, anaesthesia delivery, dialysis machines, pacemakers and ICDs, and hybrid closed-loop insulin systems — the most visible consumer-facing example of medicine automating a decision that used to be a human's.

Safety here is governed by device engineering, not software convention: IEC 60601 for electrical medical equipment, real-time constraints handled by an RTOS, and a regulatory pathway that treats the algorithm as part of the device.

The accountability question

Here's the part informatics should own rather than leave to engineers.

When a closed-loop system gives a dose that harms someone, who decided? Not the nurse — she wasn't asked. Not quite the manufacturer — the device did what it was designed to do, in a situation the design contemplated. The clinician set a target hours ago and had no view of the individual actions.

Automation didn't remove the decision. It moved it upstream and spread it thin — into a tuning parameter, a target range, a validation study, a procurement choice. That is the same structural problem as AI-based clinical decision support, arriving from a different direction and about fifty years earlier.

Control engineers solved the maths a long time ago. The governance is still open, and it is not an engineering question.