Deep Learning Model Reached 0.97 Accuracy for MACE Prediction

A new hybrid framework combined a one-dimensional convolutional neural network with a large language model to move beyond simple cardiovascular risk scoring and into patient-facing prevention guidance. The target was major adverse cardiovascular events, a category that continues to drive substantial morbidity and mortality because bedside decisions often depend on risk recognition before the first event occurs. In this design, emergency medical record data from Chungbuk National University Hospital were preprocessed, split into training and test sets, and balanced with SMOTETomek to address class imbalance before model training began.

The predictive arm centered on a 1D-CNN, a structure well suited to pattern recognition in tabular clinical features when interactions among variables may be nonlinear and difficult to capture with conventional scoring systems. In practice, that matters because cardiovascular risk rarely follows a single-axis logic. Blood pressure, medications, prior history, and other recorded features can interact in ways that are difficult to compress into one heuristic. A convolutional model can learn those feature combinations directly, which helps explain why the system reached an overall accuracy of 0.97.

For the MACE class itself, precision and recall both reached 1.00, and the F1-score was 0.98. Those values indicate that the classifier identified event cases with striking consistency in the reported test setting, while still preserving a balanced view of false positives and false negatives through the F1 metric. In a real clinical workflow, this kind of performance would be attractive for triage support, pre-visit risk stratification, or identifying patients who may merit more urgent preventive attention. The abstract does not provide calibration data, external validation, or prospective bedside testing, so the findings remain tied to the reported dataset rather than a universal clinical guarantee.

RAG Pipeline Turned Predictions Into Guideline-Grounded Advice

The second arm of the framework is where the project moves from risk estimation to clinical communication. A retrieval-augmented generation pipeline, built with prompt engineering and powered by the MISTRAL AI model, generated personalized recommendations anchored to WHO and AHA cardiovascular prevention guidelines. In plain clinical terms, the system was not asked to improvise management advice from scratch. It was constrained to retrieve and assemble guidance from established cardiovascular prevention sources, a design choice that reduces the chance of free-floating language model speculation.

This architecture fits an important gap in cardiovascular informatics. A model that predicts risk without explaining prevention leaves clinicians with a number but not a plan. Conversely, generic advice without risk stratification can miss the patients most likely to benefit from targeted counseling. By linking prediction to recommendation generation, the framework tries to close that gap in a single workflow. The reported web interface is part of the same logic: clinicians or users can receive both a risk estimate and a structured set of preventive suggestions in one place.

The recommendation system was judged with a rubric-based LLM evaluator across several dimensions: guideline fidelity, clinical safety, dietary plans, lifestyle and exercise guidance, and caution around over-the-counter medication use. The strongest scores were reported for guideline fidelity, with 5 out of 5 for both WHO and AHA alignment, and for dietary and lifestyle guidance, also rated 5 out of 5. Clinical safety received a 4 out of 5, while OTC medication caution scored 3 out of 5. Those ratings suggest the system was strongest where guidelines are relatively structured and repeatable, and less complete where safety nuance around self-medication becomes more individualized.

A one-sentence takeaway stands out. The prediction engine performed strongly, and the recommendation engine stayed close to established prevention guidance.

What the Metrics Suggest About Bedside Use

The reported accuracy is impressive, but cardiovascular clinicians will immediately ask a deeper question: what happens when the model is moved outside the dataset that trained and tested it? Emergency medical records often contain missingness, coding variability, and institution-specific documentation habits. A model that performs well after preprocessing and class balancing can still behave differently when exposed to another hospital’s workflow, another population, or another prevalence mix. The abstract does not include external validation, subgroup analysis, or prospective workflow testing, so clinical deployment would require further evidence before this could be treated as routine decision support.

Still, the biological and operational rationale is coherent. MACE prevention depends on catching risk early enough to alter behavior, optimize therapy, and reduce exposure to modifiable drivers of plaque instability and thrombotic events. An algorithm that identifies high-risk patients and pairs them with guideline-based advice could, in principle, improve preventive timing. The value would be greatest if the system helps clinicians prioritize counseling, reinforce diet and exercise recommendations, and flag patients who need caution around unsupervised OTC medication use. Those are exactly the types of bedside decisions where fragmented clinical judgment can benefit from structured decision support.

The rubric scores also reveal where the current implementation is less mature. A clinical safety score of 4 out of 5 is encouraging, but not enough to assume perfect bedside reliability. OTC caution at 3 out of 5 suggests the system may need tighter controls around self-treatment advice, since over-the-counter products can interact with cardiovascular disease states and prescription regimens in clinically meaningful ways. In a prevention workflow, that area is not cosmetic. It is often the difference between harmless advice and avoidable harm.

A Hybrid Framework With Promise, But Still a Single-Center Model

This work is best read as a technical and clinical prototype rather than a finished care pathway. Its strengths are clear: structured risk prediction, guideline-grounded recommendation generation, and a unified interface that attempts to translate analytics into action. Its limits are equally clear: a single-center EMR source, no externally validated transportability, and no direct demonstration that the model changes patient outcomes. The abstract also does not report implementation burden, clinician acceptance, or how often the generated recommendations would need editing before use.

For now, the contribution lies in showing that deep learning and large language models can be linked in a way that preserves clinical framing instead of treating prediction and advice as separate products. In cardiovascular medicine, where prevention depends on both accurate triage and credible counseling, that integrated design is the central advance. The next step will be proving that the same performance holds in broader, messier clinical settings, where risk prediction must survive variable documentation, real-world workflow pressure, and the demands of direct patient care.


Reporting by Dr. Aishwarya N., MBBS for Medicine World Foundation, with additional data provided by Frontiers in Immunology.