Why wasn't cholesterol treatment intensified? An AI note review found 88.7% of above-goal encounters stayed unchanged
In one Northern California health system, 13,810 of 28,113 patients with established cardiovascular disease had LDL cholesterol at or above 70 mg/dL. A validated language-model pathway found treatment was not intensified in 11,842 relevant notes, but missing documentation and a single-system design limit what the reasons can prove.
The short version
- What happened
- The retrospective cohort included 28,113 adults with established atherosclerotic cardiovascular disease; 13,810, or 49.1%, had LDL cholesterol at or above 70 mg/dL at their most recent encounter.
- Who it affects
- Patients with established cardiovascular disease may benefit from clearer explanations of their LDL target and why treatment is being continued, changed or deferred.
- What remains unproven
- Retrospective observational study from one Northern California academic health system; it did not test an intervention or establish causation.
Keep in mind
Retrospective observational study from one Northern California academic health system; it did not test an intervention or establish causation.
The model read the most recent encounter note, which may omit decisions distributed across clinicians, visits or systems.
Reasons for nonescalation were inherently ambiguous and produced the lowest reported task performance, with micro-weighted F1 of 0.80.
Only 40.4% of nonescalation notes contained a classified reason, so the frequency of named barriers is incomplete.
No effect on prescriptions filled, achieved LDL, cardiovascular events, patient understanding, clinician workload or cost was measured.
Multi-source analysis · 3 linked sources. Primary analysed source date: 7 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: most above-goal encounters ended without more treatment
The study found a large care gap inside a population already known to be at high cardiovascular risk. Researchers identified 28,113 adults with at least two clinician-billed encounters for established atherosclerotic cardiovascular disease between 2019 and May 2025. At the most recent encounter, 13,810 people—49.1%—had low-density lipoprotein cholesterol of at least 70 milligrams per decilitre. That threshold is commonly used to trigger closer review in secondary prevention, although the appropriate target depends on the guideline and the person's risk. The paper does not show that every value above 70 required an immediate prescription change, but it establishes the denominator for a substantial group needing an explicit decision.
In the model's inference set, 13,354 relevant notes represented 13,394 patients with above-goal LDL cholesterol. Lipid-lowering therapy was not intensified in 11,842 notes, or 88.7%. A treat-to-target strategy was documented in 3,756 notes, or 28.1%, most often a target below 70. The authors also report that clinicians who documented a target were more than three times as likely to intensify treatment. That association cannot prove that writing a target caused action: clinicians already intending to change therapy may be more likely to document a goal. It does show that explicit target-setting and treatment decisions were unevenly visible in the record.[1]
How the language-model pathway was tested
The researchers did not ask a general chatbot to summarise every chart. They built a structured extraction pathway using GPT-5 through a secure application programming interface and DSPy, an open-source framework for programming and optimising language-model workflows. The pathway looked for four things: whether a target LDL value was discussed, what target was named, whether therapy was adjusted and which reasons were documented when it was not intensified. Patients were separated into development, test and inference partitions so notes from one patient did not leak between evaluation and large-scale use.
Two senior adjudicators created the criterion-standard test annotations. The test size was planned around 1,218 notes, assuming a five-per-cent minority-class prevalence and a target confidence-interval width. On that test set, the pathway achieved an F1 score of 0.91 for whether a target was present, a micro-weighted F1 of 0.89 for the target value, 0.98 for the treatment-adjustment decision and 0.80 for reasons for nonescalation. Model–annotator reliability was not statistically different from average inter-annotator reliability. That is credible evidence for this extraction task in this dataset, not proof that the system will retain performance in other hospitals, specialties, languages or documentation styles.[1]
The distinction that matters
How the study moved from records to prevalence estimates
The language model classified documentation; it did not decide whether treatment should change.
Cohort
Identify 28,113 adults with established cardiovascular disease
13,810 had LDL at or above 70 mg/dL
Test
Compare the pathway with adjudicated notes
F1 scores ranged from 0.80 to 0.98
Infer
Apply the pathway to 13,354 relevant notes
11,842 showed no treatment escalation
Interpret
Count documented reasons
A specific reason appeared in 40.4% of nonescalation notes
What the documented reasons did—and did not—say
Only 4,788 of the 11,842 nonescalation notes—40.4%—contained at least one classified reason. Patient-reported adverse effects appeared in 1,683 notes, 12.6% of all relevant notes. LDL cholesterol being close to goal appeared in 922, patient preference to avoid intensification in 914, a preference to try lifestyle change in 811 and deferral until repeat laboratory tests in 617. These categories overlap: one note can mention more than one reason. They should be read as documented themes, not as mutually exclusive causes or percentages of all patients experiencing each barrier.
The missing-reason group is just as important. Of 7,054 nonescalation notes without a classified reason, 3,922 documented a plan to continue current therapy. The record may omit a conversation, a prior intolerance, another clinician's plan, a prescription outside the captured system or a judgment that no immediate change was appropriate. The pathway analysed the most recent relevant encounter rather than the full longitudinal decision across multiple practitioners. Absence of a reason in the note therefore does not establish absence of clinical reasoning, and an extracted reason does not tell whether the choice was shared, guideline-concordant or implemented later.[1]
Why the finding matters for patients and clinicians
For a patient who has already had coronary, cerebrovascular or peripheral vascular disease, lowering LDL cholesterol can reduce the risk of future cardiovascular events. Yet intensifying therapy is not a single mechanical step. People may have experienced muscle symptoms, worry about side effects, face insurance barriers, prefer another attempt at diet and activity or need laboratory review. A useful digital intervention must preserve those distinctions. A generic alert that says 'LDL above goal' risks repeating information clinicians already know while adding workload and ignoring why a decision was deferred.
The study points toward a more selective use of language models: summarising the documented barrier so a care team can decide what follow-up is appropriate. That might mean a medication review, a conversation about absolute benefit and harms, a repeat test with a defined date, help with cost or a referral to the clinician responsible for prevention. But no such intervention was tested. The paper measured notes, not prescriptions filled, LDL values subsequently achieved, heart attacks prevented, patient understanding or clinician burden. Turning extraction into care requires a prospective workflow with human review, error monitoring and a route for patients to correct incomplete documentation.[1]
Equity, privacy and implementation questions remain
The cohort came from one integrated academic health system in Northern California. The above-goal group included 22.2% Asian, 6.3% Black, 8.9% Hispanic or Latino and 69.6% White patients in the reported categories, and women made up 42.8%. The authors assessed model performance across key subgroups and reported no meaningful differences, but subgroup evaluation inside one system cannot settle transportability. Smaller groups may yield imprecise estimates, documentation quality can vary with access and language, and a model can reproduce gaps in what clinicians choose to write even when classification accuracy appears similar.
Clinical notes also contain highly sensitive information. The study used a secure interface approved for high-risk data, but a deployed service would need clear access controls, data-retention limits, audit logs and rules about whether information leaves the health system. Staff need to know when model-generated labels are suggestions rather than verified facts. Patients should not be categorised as refusing treatment when a note actually records uncertainty, a temporary deferral or a clinician-led decision. Privacy and contestability are therefore part of clinical validity, not administrative details added after the model is built.[1]
What evidence would change the assessment
Confidence would rise if independent health systems reproduced the extraction results with preregistered labels, transparent confusion matrices and enough patients to examine performance by race, ethnicity, sex, age, language, insurance and specialty. The pathway should be tested against a fuller longitudinal record and compared with simpler structured queries or clinician-coded reasons. Prospective trials should then randomise clinics or patients to ordinary care versus a reviewed, reason-specific intervention and measure treatment changes, LDL reduction, adverse effects, patient-reported decision quality, clinician time and cardiovascular outcomes.
The assessment would weaken if performance falls outside Stanford, if extracted reasons routinely omit shared decisions or if alerts increase prescribing without improving informed care. Funding came from the US National Heart, Lung, and Blood Institute and the Doris Duke Foundation; the paper says funders had no role in the work. One senior author disclosed consulting fees from several pharmaceutical and health-technology companies outside the submitted study. The paper is a substantial demonstration of structured note analysis, but it is not Breaking and does not justify autonomous treatment recommendations. Its strongest conclusion is that a validated model can make a hidden documentation pattern measurable enough to test better care pathways.[1][2][3]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring clinical documentation: compare the finding and its evidence limits.
Do AI scribes help equally across patient languages? A 209,009-encounter study found uneven associations
Same-day note closure was 10.2 percentage points higher with an AI scribe in non-English language-concordant visits and 1.7 points higher in English visits after adjustment. Interpreter-mediated estimates were too imprecise for a conclusion, and the study did not test note accuracy or patient outcomes.
Read the next reportExplore the full source and evidence record
What this means for people
- Patients with established cardiovascular disease may benefit from clearer explanations of their LDL target and why treatment is being continued, changed or deferred.
- Clinicians could use reviewed extraction to find follow-up needs, but an automated label must not replace shared decision-making or turn incomplete documentation into a clinical fact.
- Health-system teams need to measure whether any AI-assisted workflow improves LDL control and patient experience without increasing alert burden, inequity or privacy risk.
Global context
Cardiovascular prevention guidelines increasingly recommend lower LDL targets for people at very high risk, but medicines, insurance, monitoring and clinical responsibility differ across countries. A model trained on US English-language notes may not transfer to systems with shared records, primary-care-led prescribing or different thresholds. The underlying problem is global—turning guidelines into appropriate, individual care—but local validation and patient involvement are essential before a note-mining workflow becomes a clinical service.
What the evidence does not yet show
- Retrospective observational study from one Northern California academic health system; it did not test an intervention or establish causation.
- The model read the most recent encounter note, which may omit decisions distributed across clinicians, visits or systems.
- Reasons for nonescalation were inherently ambiguous and produced the lowest reported task performance, with micro-weighted F1 of 0.80.
- Only 40.4% of nonescalation notes contained a classified reason, so the frequency of named barriers is incomplete.
- No effect on prescriptions filled, achieved LDL, cardiovascular events, patient understanding, clinician workload or cost was measured.
What to watch next
- Independent external validation across health systems, languages and documentation cultures
- Prospective trials of reviewed, reason-specific follow-up rather than generic alerts
- Patient-reported understanding, preferences and ability to correct extracted labels
- Effects on achieved LDL, treatment harms, clinician workload and cardiovascular outcomes
- Privacy, audit and data-retention controls for language-model processing of clinical notes
Living evidence record
Impact record IAI-11846DC
Evidence stage
Studied
Confidence
Corroborated
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
8 November 2026
Source trail
3 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 8 November 2026
Research question, topics and regions
Whether a language-model pathway can reliably extract documented reasons for not intensifying lipid-lowering therapy in patients with established cardiovascular disease and above-goal LDL cholesterol
Topics: Cardiology · Clinical documentation · Large language models · Health equity · Secondary prevention
Regions: United States · North America · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.