LLM-Cardio: A Large Language Model-Based Assistant for Cardiovascular Health Inquiry and Diagnostic Support Using Wearable Data

Cardiovascular diseases (CVD) are the leading cause of death in the world, they severely affect the human life and health [1]. They are a set of conditions affecting the heart and blood vessels. The disease can be physically experienced indicating a symptomatic person or worse not feeling anything at all indicating an asymptomatic person. Most symptoms include certain types of chest pain, fatigue, palpitations, dizziness or fainting, swelling of the legs, ankles and feet and shortness of breath. To prevent patients from suffering further damage, it is essential to diagnose heart disease accurately and in a timely manner. Recently, innovative medical techniques, such as those based on artificial intelligence have been used in the medical field [1].

Among these innovations, the Internet of Wearable Things (IoWT) and large language models (LLMs) such as: CHATGPT and BERT stand out as powerful tools with the potential to reshape medical diagnostics and patient monitoring [2]. Research on LLMs based on wearable data are still in its early stages, exploring the integration of physiological data such as heart rate, sleep patterns, and physical activity into AI-driven models. However, the effective integration of IoWT real-time wearable data through connected devices with the advanced reasoning and natural language understanding capabilities that LLMs offer presents diverse opportunities to create intelligent systems capable of supporting patients’ diagnosis process which requires many tests: blood pressure, glucose, vital signs, etc., clinical studies, patient history and answers to their questions [3].

CVD remains a main concern for our wellbeing, demanding early detection, accurate risk assessment and ongoing monitoring [4]. However, the current healthcare system is often woefully inadequate in making timely, personalized diagnoses, especially for individuals in geographically disadvantaged or resource-limited settings. As widespread as wearable technology has become, its information often lies unused because there are no smart systems to filter and make sense of this information. Moreover, traditional machine learning algorithms for disease diagnosis often struggle in regards to limited generalizability, reliance on human feature engineering, and inability to handle unstructured clinical data such as medical texts.

Despite recent advances, LLM-based healthcare systems remain limited by a lack of cardiology-specific knowledge, weak integration of wearable physiological data with textual reasoning, and reliance on cloud-based deployment. These limitations hinder reliable multimodal clinical assessment and raise concerns regarding privacy and accessibility. Consequently, there is a clear need for an intelligent cardiology assistant that supports domain-adapted reasoning, multimodal data fusion, and efficient local deployment.

In this context, the objective of our work is to design and implement an intelligent cardiology assistant capable of analysing real-time, medical health data records and other relevant data, such as family history and lifestyle factors to generate reliable diagnostic insights of CVD. Our approach focused on fine-tuning an AI-powered language model on various patients cardiology-related data, including medical records, test results, and clinical observations. The system combines a locally hosted large language model, simulated wearable data, and a mobile application to create a full-stack solution for cardiac monitoring and personalized diagnostics.

To provide a comprehensive understanding of applying a LLM for cardiovascular diagnostic support and risk assessment based on wearable device data, our paper is organized as follows: Section 2 presents the background and offers an overview of LLMs in the context of wearable health monitoring systems. Section 3 details the system design. Section 4 describes the implementation. Finally, Section 5 concludes the paper and outlines potential directions for future research.

CVD is a set of conditions affecting the heart and blood vessels [13], often linked to critical symptoms such as chest pain, shortness of breath, fatigue, and irregular heartbeat, as well as risk factors including high blood pressure, cholesterol levels, obesity, smoking, and family history. The proposed approach adapts a large language model (LLM) to cardiology through a specialized dataset (symptoms, patient histories, ECG, tests, Q&A), enabling precise predictions, relevant explanations, and early diagnosis support, as shown in Figure 1. This system enhances risk assessment, clinical reasoning, and patient engagement, making digital cardiology more personalized and accessible.

3.1 Data

This work used medical data from a set of structured, unstructured and custom made cardiology related datasets. It is beneficial to explore the different data and present it in a more understandable way for the model that will be used by translating it into an accessible text format. This allows it to be interpreted in an appropriate context suitable for LLMs.

tu_pian_73.png

Figure 1. System architecture

3.1.1 Data collection

• Kaggle Cardiovascular diseases (Cardio): the dataset consists of 70 000 records ofv patient’s data, 11 features like: age, height, weight, vitals, lifestyle factors, etc. with target that indicates the presence or absence of heart disease. All of the dataset values were collected at the moment of medical examination.
• UCI Cleveland dataset: this multivariate dataset is collected from the Cleveland Clinic Foundation, consisting of 14 attributes (13 features and one target) and 303 instances, this dataset’s main use is to classify whether a patient has heart disease based on a variety of medical attributes.
• HealthCareMagic dataset: this dataset consists of 100k anonymized doctor-patient conversations, each entry comprises a patient's query and the corresponding doctor's response. This dataset is particularly valuable for training medical chatbots providing them with medical education. In our project, we retrieved the dataset directly from Hugging Face using their dataset hub for convenience and standardized access.
• Custom Cardiovascular Diseases Knowledge Base: the dataset includes 105 cardiac conditions that were collected from Texas Heart and Victor Chang cardiac institute sites. For each condition we specified its description, causes, symptoms, diagnosis, tests, treatment and prevention.
• Question and answer dataset: the dataset consists of 700 rows with question and answer pairs about CVD and general cardiology with many cardiology multiple choice questions (MCQS) that were collected from multiple cardiology books.
• Clinical cases dataset: the dataset consists of 310 clinical cardiovascular cases which were collected from different clinical books. It was constructed as a chain of thought (COT) dataset with reasoning process in disease diagnosis and its final answer [14, 15].

3.1.2 Data pre-processing

In this initial phase, data from multiple sources is consolidated and cleaned to create a unified dataset. This preparation is essential for aligning the data into a format suitable for detailed analysis and subsequent processing.

5. Data cleaning

• We checked for duplicates and missing values in both Cleveland and Cardio datasets.
• Removed all rows containing missing or null values in Cleveland dataset to ensure data quality and avoid bias during training.

5. Feature engineering

This is the process of transforming raw data into meaningful inputs by creating, modifying, or scaling features to improve model performance.

• BMI calculation: We created a new feature for body mass index (BMI: see Table 2) BMI provides a reliable indicator of body fat for most people. Therefore, it is used to screen for weight problems that may lead to health concerns. Using the formula [16]:

$\mathbf{BMI}=\frac{\text{Weight}\left( \text{Kg} \right)}{\text{Height}{{\left( m \right)}^{2}}}$

• Age transformation: We converted the age column from days to years by dividing by 365.

Table 2. Body mass index [16]

5. Target variable scaling

We converted target variable for Cleveland dataset to:

• Value 0 = no disease.
• Values 1–4 = disease present.

5. Data filtering

We applied this method to Healthcare Magic dataset selecting only records related to cardiology and cardiovascular diseases using cardiology keywords matching like: arrhythmia, angina and pacemaker, etc. We ended with 19 253 records.

5. Data balancing

• Target Class Balancing: From the original dataset of 70 000 rows, we sampled 10 000 rows to ensure a 50/50 balance between patients with and without heart disease, as shown in Figure 2.
• Gender Balancing: Further refined the sampled data to ensure equal representation of genders within each target class for fair model training.

tu_pian_74.png

Figure 2. Balanced cardiovascular dataframe

5. Text format conversion

To prepare the UCI Cleveland and Kaggle Cardio datasets for language model fine tuning, each record was transformed from tabular format into descriptive text using natural language templates. These representations describe patient attributes, medical findings, and diagnostic outcomes in a format suitable for instruction-based learning.

Natural language templates were designed to include all clinically relevant attributes in a coherent and medically interpretable manner. The process relied exclusively on observed data and was implemented deterministically, with automated controls ensuring completeness and reproducibility.

5. Instruction formatting

All datasets were converted into an Instruction, Input, Output (IIO) format to enable instruction tuning, a special case of supervised fine-tuning of a language model where each training example includes a task instruction along with input and output, as illustrated in Figure 3. The goal is to make the model better at following human instructions.

For each example:

• The Instruction specifies the task the model should preform, such as: diagnosis, explanation or prediction.
• The Input contains the context or relevant information, like: patient symptoms, history or question.

The Output contains the desired model response like: diagnosis, answer, or explanation.

tu_pian_75.jpg

Figure 3. Example of instruction input output data format

5. Merging all datasets

All pre-processed datasets were unified into a single dataset as a JSONL format (Cardiac_10sft) ensuring the instruction format for fine-tuning. The jsonl format is ideal for large datasets where each line is a separate JSON object and compatible with any large language model training tools like LoRA and PEFT.

5. Prompt template

The final merged dataset was formatted using a custom prompt template that ensures clinical safety, structured reasoning, and multilingual response capability, as shown in Figure 4. Its main purpose in fine-tuning is to teach the model what the task is, ensure it understands the context and can generate a consistent task specific responses especially important for medical domain.

tu_pian_76.jpg

Figure 4. Train dataset prompt template

5. Data splitting

The dataset had 30 720 rows. By randomly splitting it into training and validation sets, with 90% for training (27, 648 rows) and 10% for validation (3072 rows). The validation data is used to evaluate the model’s performance on unseen data during fine-tuning to monitor if the model is learning effectively without overfitting.

3.2 Model

LLMs have shown great success in medical natural language processing (NLP) tasks when fine-tuned on domain-specific instructions, making them suitable for interpreting medical symptoms, generating diagnoses, and responding to health related questions in human-like language [17].

3.2.1 Llama-3.1 large language models for text generation

The Meta-Llama-3.1 collection of multilingual LLMs is a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes (text in/text out). The Llama-3.1 instruction tuned text only models are optimized for multilingual dialogue use cases and outperform many of the available open-source and closed chat models on common industry benchmarks [18].

3.2.2 Model selection: Meta-Llama-3.1-8B-4bit

The pre-trained model used as a base is the Meta-Llama-3.1-8B-Instruct variant, released and optimized by the Unsloth project for 4-bit quantized inference and training. This model was chosen for its strong instruction following capabilities, compact size relative to performance, compatibility with quantization and LoRA fine-tuning and its strong performance in understanding complex natural language queries, open-source accessibility.

3.2.3 Instruction tuning for medical task adaptation

To adapt the model to the cardiology domain, we performed Instruction Tuning, a form of Supervised Fine-Tuning (SFT). This involves training the model on curated examples formatted as instruction–input–output triples, covering a wide range of tasks such as:

• Symptom based diagnosis.
• Medical explanation generation.
• Disease diagnostic support and risk assessment.
• Cardiology-related questions and answers.
• Interpretation of clinical test data.

3.2.4 Fine-tuning approach

To efficiently adapt a pre-trained large language model (LLM) to our task of cardiac disease diagnostic support and risk assessment, we employed a parameter-efficient fine-tuning (PEFT) strategy, specifically the LoRA method, in conjunction with modern libraries and tools (Transformers, Unsloth, PEFT, Hugging Face Hub, LoRA, Runpod, TRL, Weights and biases) that support optimized training on limited hardware.

5. LoRA and Training Configuration

Training process followed these hyperparameters: total batch size of 2 per device with gradient accumulation steps = 4 so the total batch size is 8, learning rate of 2 × 10-4, 2 epochs, maximum sequence length of 4096 tokens, and a warmup rate of 0.01 with weight decay of 0.01. Optimizer: AdamW (8-bit). Quantization: 4-bit.

• Rank r: 8
• LoRA-Alpha: 16

The LoRA hyperparameters (r = 8, LoRA-Alpha = 16) were selected based on empirical evaluation and PEFT literature. This configuration provides an optimal trade-off between model expressiveness and computational efficiency under 4-bit quantization.

5. Training and Validation Results

The model demonstrated good convergence during training.

• Training Loss: showed a consistent and steady decline, starting around 1.1 and decreasing to below 0.9 by the end of the second epoch, indicating effective learning without divergence, as shown in Figure 5.
• Validation Loss: decreased from 1.04 after the first epoch to approximately 0.94 after the second, showing continued generalization improvement and no signs of overfitting. The trend suggested that the model might still benefit from an additional training epoch, as shown in Figure 6.
• Training Time: 4 hours and 47 minutes.

tu_pian_78.jpg

Figure 5. Train loss plot

tu_pian_79.jpg

Figure 6. Validation loss plot

3.2.5 Evaluation

BERTScore is an automatic evaluation metric for text generation that computes a similarity score for each token in the candidate sentence with each token in the reference sentence. It leverages the pre-trained contextual embeddings from BERT models and matches words in candidate and reference sentences by cosine similarity. Moreover, BERTScore computes precision, recall, and F1 measure, which can be useful for evaluating different language generation tasks [19].

${{R}_{\text{BERT}}}=\frac{1}{\left| x \right|}\underset{{{x}_{i}}\in x}{\mathop \sum }\,\underset{{{{\hat{x}}}_{j}}\in \hat{x}}{\mathop{max}}\,\mathbf{x}_{i}^{\top }{{\mathbf{\hat{x}}}_{j}}$

1. Token Matching via Cosine Similarity: For each candidate token, compute cosine similarity with all reference tokens.
2. Precision: Measures how many tokens in the candidate have a similar counterpart in the reference.
3. Recall: Measures how many tokens in the reference have a match in the candidate.
4. F1-score: The F1-score is the harmonic mean of precision and recall; It provides a single summary value of overall semantic alignment between the candidate and the reference.

${{F}_{\text{BERT}}}=2\frac{{{P}_{\text{BERT}}}\cdot {{R}_{\text{BERT}}}}{{{P}_{\text{BERT}}}+{{R}_{\text{BERT}}}}$

We evaluated our fine-tuned model using a test dataset structured in an instruction–response format. The records combined question–answer pairs, CVD diagnostic support and risk assessment data derived from the cardio dataset, as well as 25 multiple-choice questions drawn from the American Nurses Association (ANA) cardiac question bank. Initially, we trained two machine learning models on our training dataset; however, their accuracy was near zero / very low due to their inability to process long text sequences. We subsequently trained these models on the pre-processed Cardio dataset (10,000 records), compared their performances, and then evaluated them on the Cardio test classification records, as shown in Figure 7. The comparative evaluation using BERTScore metrics is summarized in Table 3.

Table 3. Model evaluation comparison

Traditional machine learning models (Naïve Bayes and Logistic Regression) are included in the evaluation solely as non-generative reference baselines to illustrate the limitations of discriminative classifiers when applied to long-form medical instruction-following tasks. Their BERTScore results are not intended to represent competitive generative performance but to contextualize the necessity of generative language models for conversational diagnostic reasoning.

Example responses

tu_pian_80.png

Figure 7. Comparison of base and fine-tuned models responses

The low BERTScore values observed for traditional machine learning models reflect their inability to generate free-form text. In contrast, LLM-Cardio demonstrates stronger semantic coherence and medical accuracy compared to BioGPT and the base LLaMA-3.1 model.