Comorbidity-aware grading, copy-forward detection and regimen profiles for attribution

What shipped on 9 March 2026

Burna AI's comorbidity-aware grading, released on 9 March 2026, puts a patient's known conditions in front of the grading engine, so an adverse event that overlaps a pre-existing condition carries a note saying so. The same release added copy-forward detection, a review queue for terms the engine cannot match, and pre-built oncology regimen profiles for multi-drug attribution.

It is written for the coordinators and investigators who grade and attribute adverse events in combination-therapy trials, where context decides what a finding means. Peripheral neuropathy in a patient with long-standing type 2 diabetes is a different question from the same finding in a patient with no neurological history.

How comorbidity-aware grading separates a new event from a known condition

Each patient's known comorbidities now enter the grading context alongside the clinical note. When a finding overlaps one of them, the graded event carries a comorbidity note: a short explanation of the overlap, so the person attributing the event sees the pre-existing condition next to the evidence instead of hunting for it in prior visits.

Two details matter in practice:

• It switches itself off when there is nothing to weigh. A patient with no comorbidities on record is graded exactly as before.
• It informs attribution; it does not decide it. The note is context for the investigator, who makes the call. AI suggests, clinicians decide.

The distinction is the same one a trial draws between a treatment-emergent event and the baseline a patient arrives with; treatment-emergent adverse events and grading against the baseline covers it in depth, and how Burna AI distinguishes a new event from a pre-existing condition answers it briefly.

Since then (May 2026): comorbidities moved onto the patient record as a list clinicians can view, add to and remove entries from.

What copy-forward detection flags

Copy-forward is the habit of carrying text from a prior note into the current one, a practice linked to note bloat, internal inconsistencies and error propagation in the record (Tsou et al., Applied Clinical Informatics, 2017). For adverse events it shows up as the same description and the same grade repeating visit after visit, supported by the same sentence, whether or not anyone looked again. A grade copied forward can hide an event that has worsened or resolved, and it leaves a monitor no evidence that the event was re-assessed.

Copy-forward detection runs after each grading run and keeps a per-patient record of what was graded at every visit. It looks for four signals:

1. The grade has not changed across visits.
2. The supporting evidence repeats almost word for word.
3. The timeline does not make clinical sense.
4. Nothing in the note shows the event was re-assessed.

A finding needs at least two signals before it is flagged, which keeps one coincidence from raising an alert. Each flag comes with a suggested next step: reassess the grade, check with the patient, review the prior notes, or mark the finding as likely valid. How Burna AI catches stale clinical documentation has the short version.

Since then (May 2026): a second version added a fifth signal, a gap between what the patient reports and what the clinician graded, and a confidence floor that holds back low-confidence alerts.

Self-improving terminology: unmatched terms go to a review queue

Clinical vocabulary varies by site. When the grading engine meets a term it cannot match to a CTCAE v6.0 term, the term lands in a review queue for that organization, de-duplicated, with a count of how often it has appeared.

A reviewer resolves each entry in one of four ways:

• Add as a synonym of a named CTCAE term
• Extraction error, when the engine pulled the wrong span from the note
• Out of scope, when the phrase is not an adverse event
• Needs investigation, when the answer is not yet clear

The queue shows exactly which phrasings the synonym map is missing, and each resolution is the input to the next synonym map update. That loop is what self-improving terminology means here: the more notes the engine reads, the more precisely the queue shows what to add.

Oncology regimen profiles for multi-drug attribution

In a combination trial, which drug caused an event is the question that shapes the label, and a note often names only the regimen. Pre-built oncology regimen profiles close that gap. They cover lymphoma, breast, gastrointestinal, lung, genitourinary, head and neck, sarcoma, myeloma, leukemia and ovarian cancer regimens, plus CAR-T conditioning: R-CHOP, FOLFOX, FOLFIRI, FOLFIRINOX, AC-T, carboplatin and paclitaxel, gemcitabine and cisplatin, and VRd among them. Which regimens have profiles lists more.

Each profile expands the regimen into its component drugs, records the role of each (cytotoxic, targeted, immunotherapy, supportive or modulator) and lists each drug's characteristic adverse events. Attribution can then score every drug in the regimen, even when the note says only "cycle 3 of FOLFOX".

Two changes in the same release feed attribution better inputs:

• Timing from the note. Grading now extracts when an event started, how long it lasted and its pattern, while it has the full note in view, and passes those fields to attribution.
• One name per drug. Brand names resolve to generic names before matching, so "Cytoxan" reaches attribution as cyclophosphamide.

Wider symptom groupings behind the engine

The hand-curated symptom groupings that sit between an organ system class and an individual term grew from 79 to 92, and from 17 organ system classes to 20. The new classes are hepatobiliary, ear and labyrinth, and reproductive system, with more gastrointestinal and infection groupings besides. Grouping lets the engine recognize that differently worded findings belong together before it chooses a CTCAE term.

To see comorbidity context and regimen-aware attribution on your own encounters, book a 15-minute conversation.