CLAUDE: IS AN ICD, AI?
An implantable cardioverter-defibrillator (ICD) includes a pacemaker function, but it also continuously monitors the heart rhythm, converting it in real time into internal parameters using advanced, flexible, adaptive software. When certain thresholds are exceeded, the device delivers a cardioversion shock. Is this autonomous AI?
Yes — read literally against the 2027 definitions, the ICD is autonomous, and specifically Level III. Walk it through the spine: the output is clinically useful (it treats arrhythmia, squarely inside the §201(h) verbs); the software converts the raw electrogram into derived parameters (rate-zone assignment, interval stability, onset, morphology-match scores, SVT discriminators) that are categorical/quantitative and well beyond mere summation, so it clears the assistive boundary; and it doesn't stop at a parameter — it reaches a conclusion ("this is a shockable rhythm") and initiates the intervention itself. That conclusion-plus-action is the autonomous trigger. Then the level: it automatically initiates a management action without a human in the loop and keeps operating unless a physician intervenes by reprogramming or deactivating. The 2027 Level III text — automatically initiates management, requires physician oversight and review of performance, continues unless a physician chooses to intervene — reads almost like it was written from a device-clinic follow-up note. Even the new transparency requirement is satisfied: the ICD logs and reports its derived parameters and episode electrograms on interrogation.
- And that is precisely the embarrassment that proves your opening complaint. Nobody calls an ICD "AI." It's a deterministic, decades-old, FDA-regulated device.
But because the taxonomy pointedly refuses to define AI and the criteria never read AI-ness, the taxonomy has no mechanism to exclude it. A 1985 rate-only defibrillator would clear the same gates. So taken at its word, the taxonomy reclassifies the entire installed base of automatic implantable defibrillators, rate-adaptive pacemakers, and closed-loop insulin pumps as "autonomous AI, Level III." When your AI taxonomy swallows the whole history of automatic implantable therapy, the word "AI" in the title is doing no work — which is exactly what you said in your first message, now with a pacemaker in it.
Two refinements worth keeping. First, the only thing that actually keeps the ICD out is the application-stage trigger — the taxonomy is invoked by code-change applications "describing use of AI-enabled medical services," and nobody files an AI-software CCA for an ICD's detection algorithm. The device is paid through implantation, interrogation, and DRG/pass-through pathways; CPT codes the procedure, not the algorithm. So the undefined term "AI-enabled" does the covert gatekeeping that the classification criteria can't — your residual incoherence, made concrete.
Second, the sophistication inverts the cleanliness. A modern discriminator-laden ICD derives parameters and is unambiguously Level III. But a primitive rate-only ICD detects without really deriving a parameter, yet still auto-acts — which drops it into the off-diagonal empty cell I flagged earlier (detection-without-a-parameter wired to autonomous action), a combination the taxonomy has no box for. So the cruder the device, the more it falls into the taxonomy's blind spot; the more advanced, the more cleanly it's captured. That's backwards from what an "AI" taxonomy should do.
#
#
Chat GPT - Same Question
#
#
Under Appendix S as written, your implantable cardioverter-defibrillator example strongly exposes the taxonomy’s problem.
My best answer: functionally, it behaves like autonomous Level III software, but Appendix S may not clearly classify it as “autonomous AI” unless the AMA first agrees that the service is within the AI/software-output taxonomy.
Here is why.
An implantable cardioverter-defibrillator does several things that match the autonomous side of Appendix S:
It continuously monitors physiologic input.
It derives internal parameters from rhythm data.
It applies adaptive software rules to determine whether a dangerous rhythm threshold has been crossed.
It initiates a therapeutic intervention—a cardioversion or defibrillation shock—without waiting for a physician to approve it in real time.
That looks very close to autonomous Level III in the 2027 Appendix S framework, where software automatically initiates management actions based on interpretations or conclusions drawn from derived parameters, and those actions continue unless a physician intervenes. The 2027 text defines autonomous software as deriving parameters and independently generating clinically meaningful interpretations or conclusions, with Level III automatically initiating management actions subject to physician oversight.
But this example also shows why Appendix S is under-engineered. The taxonomy is framed around “AI-enabled” services, while also saying that “AI” is not defined. An ICD may use sophisticated software, adaptive thresholds, signal processing, and decision logic, but many people would call it an advanced implanted device rather than “AI.” Others could plausibly say it is algorithmic autonomous clinical software. Appendix S does not give a gatekeeping rule to settle that dispute.
So the answer depends on the question being asked.
If the question is, “Does an ICD behave like autonomous software under the Appendix S control model?” then yes, very much so. It detects/derives rhythm parameters, reaches a clinically meaningful conclusion, and initiates therapy without concurrent physician involvement.
If the question is, “Would AMA necessarily classify an ICD as autonomous AI under Appendix S?” then not necessarily. The taxonomy does not define AI, does not provide a flow chart, and does not explain whether long-established embedded device algorithms are inside or outside the AI category.
This is a perfect stress-test example. It separates method label from clinical behavior. By behavior, the ICD is autonomous. By branding or regulatory convention, it may not be called AI. That is exactly the paradox: Appendix S classifies output roles better than it defines the universe of things to which the classification applies.
#
#
One might also argue, AMA CPT provides codes for the surgical procedure of implanting an ICD. The CPT doesn't really care if the surgeon is implanting an ICD or a teacup. Of course, it's specified as ICD to allow surveys and minutes of RVU time and so on. But it's a surgical procedure this morning that is coded, not a distant future electronic event.
#
Yes — that is exactly the counterargument, and it is a strong one.
The ICD example only becomes an “autonomous AI” puzzle if we treat the implanted device’s future behavior as the service being classified. But CPT may be doing something much narrower. It is coding the physician service performed today: implanting, replacing, programming, interrogating, or managing an ICD. The code descriptor names the ICD because the device type affects physician work, surgical complexity, survey valuation, practice expense, risk, and clinical context. But CPT is not necessarily trying to classify every future algorithmic event the device may perform after implantation.
So the teacup analogy is useful, with a caveat. CPT does care that it is an ICD rather than a teacup because the surgical work and clinical purpose are different. But the unit of coding is still the medical service, not the autonomous downstream life of the object. A surgeon implants an ICD in the morning. Months later, the device detects ventricular tachycardia and delivers a shock. That future shock is not the surgeon’s separately coded morning procedure.
This helps distinguish two things that Appendix S may blur:
1. A service that uses software output as part of the coded service.
Example: retinal imaging software analyzes an image and produces a diagnostic or referral output that is itself central to the billed service.
2. A device whose implanted function includes autonomous therapeutic behavior.
Example: an ICD that continuously monitors rhythm and later delivers therapy, but where CPT may be coding implantation, interrogation, programming, or device management—not each future internal decision event.
That distinction weakens the argument that ICD implantation itself must be “autonomous AI.” The implanted ICD may contain autonomous control logic, but the CPT code for implantation is not necessarily a code for the autonomous logic. It is a code for placing the device.
This is another reason a logic tree would help. It should begin with a gate question such as:
What is the CPT service being classified?
Is the software output itself part of the billed medical service, or is the software merely a capability of a device that is implanted, maintained, or monitored under separate procedural codes?
Without that gate, Appendix S risks confusing the coded physician service with the future behavior of an algorithmic device. That is not a small distinction. It is central to how CPT works.
