Medical History Review in Clinical Dental AI: Flagging Anticoagulants, Bisphosphonates and ASA Status Before the Appointment

How many of tomorrow’s patients take an anticoagulant, and when did anyone on your team last check the dose against what was actually prescribed? In most practices, the answer comes out at the chair — after the patient is seated, the operatory is set up, and the hygienist is already reading the health history aloud.

That said, the information usually exists before the appointment. It sits in a signed intake form, a medication list in the practice management system, a scanned referral letter, or a free-text note from the last visit that nobody re-read.

Medical history review before the appointment is a high-value, low-autonomy task for a clinical AI agent. Below, we cover what an agent can safely surface — anticoagulants, antiplatelets, bisphosphonates and other antiresorptives, and ASA physical status — and where the clinician’s judgment must stay in the loop.

Why Pre-Appointment Flagging Breaks Down Today

Keep in mind that in most practices, health history review happens at the point of care. The provider skims the form during setup, asks whether anything has changed, and initials the update — a process that works until the schedule runs forty minutes behind.

The ways it fails are predictable, and they include but are not limited to:

• Stale lists. The patient started apixaban after a cardiology visit in March, but the dental record still shows the medication list from two years ago.
• Everyday wording. Patients write “blood thinner,” “baby aspirin,” or “bone pill,” and nobody maps those phrases to a drug, a dose, or a risk class.
• Buried indications. A history of breast cancer appears on page two, while the zoledronic acid infusions that matter for an extraction are never written down at all.
• Late discovery. The flag comes up with the patient in the chair, when the only safe option for a planned extraction is to reschedule and lose the slot.

All of these are retrieval and reconciliation problems rather than clinical judgment problems. That difference is the whole design brief for the agent.

What A Clinical AI Agent Can Safely Surface

The safe scope is narrow. The agent reads every source in the chart, pulls out medications and conditions, standardizes them, and matches them to risk classes the practice has already defined; accordingly, its output is a flag, the evidence behind it, and a list of what is missing.

Anticoagulants And Antiplatelets

Warfarin, the direct oral anticoagulants (apixaban, rivaroxaban, dabigatran, edoxaban), and antiplatelets such as clopidogrel, ticagrelor, prasugrel and low-dose aspirin all change bleeding risk for extractions, periodontal surgery, and implant placement. What’s more, dual antiplatelet therapy after a recent stent, or an anticoagulant combined with an antiplatelet, raises that risk further.

For warfarin, the flag should carry the most recent INR value and its date. The Scottish Dental Clinical Effectiveness Programme’s guidance, for instance, generally supports routine extractions in a stable patient whose INR is below 4 and was checked ideally within 24 hours of treatment — so an INR from six weeks ago is a flag in itself.

For DOACs, some guidance suggests changing the timing of the morning dose before procedures with higher bleeding risk. That is exactly the decision the agent must not make, so it surfaces the drug, the dose, the prescriber, and the planned procedure and lets the dentist decide.

Bisphosphonates And Other Antiresorptives

Alendronate, risedronate, ibandronate and zoledronic acid raise the risk of medication-related osteonecrosis of the jaw (MRONJ) after extractions and other invasive procedures, and so does denosumab, which is an antiresorptive but not a bisphosphonate. Antiangiogenic cancer drugs such as bevacizumab and sunitinib belong on the same watch list.

Indication and route matter more than the drug name. The American Association of Oral and Maxillofacial Surgeons’ 2022 position paper puts MRONJ risk for osteoporosis patients on oral bisphosphonates at a small fraction of one percent, while cancer patients on high-dose IV zoledronic acid or denosumab face risks in the low single-digit percentages.

As a result, a useful flag reads “zoledronic acid, IV, oncology dosing, last infusion unknown” rather than simply “bisphosphonate.” Duration matters as well, since longer total use of oral bisphosphonates — generally beyond about four years — is linked to higher risk.

ASA Physical Status

The American Society of Anesthesiologists’ physical status classification runs from ASA I (a normal healthy patient) to ASA VI (a declared brain-dead organ donor), with an E modifier for emergencies. In dental settings, the line between ASA II and ASA III often decides whether sedation happens in the office, whether a medical consult is required, and how long the appointment should be.

An agent can gather the evidence behind the class: diagnoses, medications that suggest how severe a disease is, recent hospital stays, and functional notes such as “short of breath climbing stairs.” It can then suggest a provisional class, citing the source of each supporting fact.

However, ASA status depends on how well a disease is controlled, which a form rarely shows. Even anesthesiologists only moderately agree when they assign ASA classes to the same patient, so the agent’s output is a draft for the clinician and never a final value written to the chart.

The Rest Of The Watch List

Most practices apply the same approach to a few other risk classes. Some examples include:

• Antibiotic prophylaxis triggers. These include prosthetic heart valves, prior infective endocarditis, and certain congenital heart conditions under the American Heart Association’s 2021 scientific statement. Prosthetic joints go to the practice’s own policy, since the ADA generally does not recommend routine prophylaxis for them.
• Vasoconstrictor cautions. Nonselective beta blockers, tricyclic antidepressants, and poorly controlled hypertension may affect the epinephrine dose in a local anesthetic.
• Immunosuppression. Chemotherapy, long-term corticosteroids, and biologics change healing and infection risk after surgery.
• Allergies. Latex, penicillin, and local anesthetic reactions should be recorded with the type of reaction, so a rash is not treated the same as anaphylaxis.

Each class should come from a list that the clinical team owns and approves. The agent applies that list and never makes up its own.

Where The Clinician’s Judgment Stays In The Loop

The boundary is simple to state and harder to enforce. The agent informs and the clinician decides, and every flag should make it obvious that a person made the decision.

Here’s how the work divides between the agent and the clinical team:

Note that the last row matters more than it looks. A message that says “please bring your latest INR result” is administrative, while one that says “skip your morning Eliquis” is a medication instruction that no agent should send.

How The Pipeline Should Be Built

Compliance comes before model quality. Medical history is protected health information, so every component — the model endpoint, the data store, the logging layer — must be covered by a signed BAA, and access should follow the minimum-necessary standard described in our guide to HIPAA-compliant clinical AI for dental practices.

With that in place, a production design typically runs in five stages:

1. Ingest. Pull the medication, problem, and allergy lists through the Open Dental API or your PMS equivalent, along with intake forms and scanned referral letters. Limit this to patients on the next one to three days of the schedule.
2. Extract. Use a model such as Claude on Amazon Bedrock, running inside your AWS account under the AWS BAA, to pull drug names, doses, frequencies, routes, and conditions from free text. This is the same approach covered in our post on Bedrock for clinical AI.
3. Standardize. Map every extracted drug to an RxNorm concept so that “Eliquis 5 mg BID,” “apixaban,” and “my blood thinner (Eliquis)” all resolve to one ingredient and one class.
4. Classify with fixed rules. Apply the practice’s approved risk-class table with a plain lookup instead of a model call. That way the same drug always produces the same flag, and every change to the table is versioned.
5. Send for review. Write flags to a review queue for the hygienist or front-desk lead. Confirmed flags then appear on the provider’s morning huddle sheet.

Keep in mind that the model’s job ends at step two. Keeping classification out of the model is what makes the system auditable, because a reviewer can trace any flag back to a source phrase, an RxNorm code, and a versioned rule.

Treat Patient-Entered Text As Untrusted

Intake forms are free text written by the public, and free text can contain instructions aimed at the model. For this reason, the extraction prompt should treat form content strictly as data and check outputs against a fixed schema — an approach we detail in our post on prompt injection in clinical AI.

Log Everything The Reviewer Saw

Every flag should record the source document, the extracted text, the model and prompt version, the rule-table version, and who confirmed or dismissed it. With that audit trail, you can answer a dental board inquiry — or a malpractice carrier — about what the practice knew before the appointment.

Why Missing Data Is The Most Useful Flag

In practice, the agent’s most useful output is often a gap rather than a finding. A DOAC with no prescriber listed, warfarin with no INR in the last week, or a cancer history with no treatment details are all problems the front desk can fix with a phone call two days ahead.

At the chair, the same gaps cost the appointment slot. Therefore, the review queue should sort gaps by appointment time and procedure type, so an extraction tomorrow morning comes ahead of a recall cleaning next Thursday.

Stale histories need their own handling. If the last signed update is older than the practice’s review window, the agent should mark every flag based on it as unconfirmed instead of giving it the same weight as yesterday’s intake form.

Groups whose records are split across acquired offices should merge duplicate charts first, as covered in our post on dental chart consolidation. Otherwise, the agent can match one patient’s history to another patient’s appointment.

How Do You Know The Flagging Is Working?

Overall accuracy is the wrong headline metric, because most patients on any schedule have no high-risk flag at all. What matters is recall — the share of real cases the agent catches — on the classes that change treatment, starting with anticoagulants, antiplatelets, and antiresorptives.

A credible test suite includes:

• A labeled history set. This means several hundred de-identified or consented real histories labeled by clinicians, including scanned forms and messy everyday wording.
• Recall and precision for each class. Report these separately for each risk class, and require a set recall level on anticoagulants and antiresorptives before any release.
• Alert burden. Count false positives per day per provider, since a huddle sheet with fifteen flags on it gets skimmed just like the paper form did.
• Shadow mode. Run the agent alongside the existing process for several weeks and compare its flags with what clinicians caught at the chair before any flag reaches a provider.

After launch, the same suite runs on every model-version change and every rule-table edit. Our posts on clinical AI evals and model drift monitoring cover the test setup in more depth.

Confirmed flags also feed later work. For instance, AI-assisted informed consent documents can name the patient’s specific bleeding or MRONJ risk instead of listing every possible risk.

Frequently Asked Questions

Here are several questions that come up when practices scope this kind of agent. Each answer assumes the clinician-in-the-loop boundary described above.

Scoping Your First Medical History Agent

Pre-appointment flagging pays for itself through fewer reschedules and conversations that happen two days early instead of at the chair. That said, it only works when the scope stays narrow and the clinician remains the decision-maker.

If you’re scoping a medical history review agent for a multi-location group, the NexV team builds and runs HIPAA-grade clinical AI on Bedrock against Open Dental and other practice management systems. Reach out for a working session — we’ll map your intake sources, define the risk-class table with your clinical leads, and leave you with a test plan your board and compliance team can sign off on.