Large Language Models Enter Cardiovascular Prevention With Clear Utility and Hard Limits
Large language models are moving into preventive cardiology faster than the evidence base can comfortably support. In this narrative review, Ferreira Santos and Dores mapped their potential across cardiovascular prevention and drew a firm boundary around where current performance still falls short of bedside needs. The central message is direct: these systems can improve access, drafting, summarization, and patient-facing education, but they are not ready to function as autonomous clinical decision-makers.
The review spans literature published from January 2015 through November 2025 and organizes the field into three domains: patient-facing tools, clinician-facing tools, and system-level tools. Each domain exposes a different kind of value proposition. Patient interactions benefit from immediate language generation and the ability to translate dense preventive guidance into plain speech. Clinicians gain speed when summarizing notes or drafting documentation. Health systems may use these models to extract structured data, build registries, and support surveillance work that is typically labor-intensive. A single limitation runs through all three: the model’s fluency can look like competence even when the underlying output is incomplete or wrong.
Patient Education Gains Reach, but Precision Remains Fragile
For patients, the attraction is obvious. A model can answer questions at any hour, use a conversational tone, and produce guidance that sounds empathetic rather than bureaucratic. In preventive cardiology, where adherence hinges on diet, exercise, blood pressure control, lipid management, smoking cessation, and long-term follow-up, plain-language reinforcement has real value. The review found that LLMs can generate guideline-concordant education and may help close health literacy gaps that often undermine prevention efforts.
The problem emerges when that same fluency is asked to stand in for individualized judgment. Preventive cardiology depends on risk context: age, comorbidities, medication burden, baseline risk, and competing priorities. An LLM can describe what general prevention looks like, but unsupervised advice becomes unreliable when the question requires nuance. It may overstate certainty, flatten nuance in shared decision-making, or deliver a response that sounds personalized without actually being anchored to the patient’s data. In prevention, a confident answer that misses the clinical context can be worse than no answer at all.
Hallucination risk is not an abstract software concern here. It can translate into wrong lifestyle advice, mistaken medication framing, or false reassurance about risk. The authors treat this as a governance issue, not a technical footnote, because preventive care often starts with information rather than procedures. When information is wrong, harm begins long before any prescription is written.
Clinician Workflow Support Is Real, but Risk Prediction Needs Guardrails
On the clinician side, the strongest near-term use case is documentation support. LLMs can summarize clinical notes, draft letters, and reduce clerical drag in busy workflows. That matters in preventive cardiology, where visits often absorb a wide range of tasks: interpreting biomarkers, reconciling medications, reviewing family history, and translating guideline targets into practical plans. Anything that reduces documentation burden has the potential to return time to patient counseling.
Still, the review draws a sharp line between language processing and deterministic prediction. These systems are not reliable enough for autonomous risk calculations or for making treatment decisions without oversight. Preventive cardiology depends on calibrated estimates of risk, thresholds for therapy, and consistent interpretation of changing evidence. An LLM may appear to reason through those tasks, but its outputs can shift with prompt phrasing, omit critical factors, or reflect outdated guidance. Temporal obsolescence is especially important in prevention, where guideline updates can change treatment thresholds and risk framing.
This is where automation bias becomes clinically dangerous. When a tool sounds authoritative, clinicians may be more likely to trust it than they should. A concise summary can improve efficiency, but a polished answer can also obscure error. The governance question is not whether these tools can write fluently; they already can. The issue is whether they can be embedded in workflows in a way that preserves clinical accountability and forces verification before action.
System-Level Uses May Expand Prevention Infrastructure
The strongest system-facing promise lies in data work. Preventive cardiology generates large volumes of information scattered across clinic notes, imaging reports, laboratory data, and patient narratives. LLMs may help extract phenotype data, support registry construction, and assist quality surveillance. Those functions are less visible than patient chat or note drafting, yet they may ultimately matter more for population health because they help turn unstructured records into analyzable information.
The review also points to multimodal risk prediction as an emerging use case. In principle, language models can help integrate different data streams that usually sit apart from one another. The appeal is substantial: better identification of patients at risk, more efficient surveillance, and potentially more targeted prevention strategies. Even here, caution dominates. Performance in silico does not guarantee safety in live clinical environments, especially when data quality is uneven, missingness is common, or model outputs feed directly into downstream decisions.
Privacy introduces another hard constraint. Systems that ingest large clinical datasets raise concerns about data governance, access control, and secondary use. In prevention, where long-term follow-up and broad population data are central, those concerns are not peripheral. They sit at the center of implementation.
C.A.R.D.I.O. as a Governance Framework for Safe Translation
To move from promise to practice, the authors propose the C.A.R.D.I.O. framework: Clinical validation, Auditability, Risk stratification, Data privacy, Integration, and Ongoing vigilance. Each element addresses a different failure mode. Clinical validation asks whether the model works in the relevant population and task. Auditability requires traceable outputs and review pathways. Risk stratification distinguishes low-risk support functions from high-stakes clinical decision points. Data privacy forces the field to confront stewardship and consent. Integration focuses on whether the tool fits real clinical workflows rather than interrupting them. Ongoing vigilance recognizes that model behavior can drift over time as data, guidelines, and use patterns change.
One sentence captures the practical reality: these systems should be treated as supervised reasoning engines, not independent clinicians. That framing matches the evidence summarized in the review. LLMs can assist with communication, documentation, and data handling, but they should not be allowed to substitute for judgment in preventive cardiology.
The review does not claim that artificial intelligence has solved prevention. It instead defines the boundaries of responsible use in a field where misclassification, overconfidence, and outdated advice can carry long tails of harm. The unanswered operational question is how quickly health systems can build validation, monitoring, and accountability into deployment before these tools become embedded by default. Until that infrastructure exists, supervised use remains the only defensible path.