| --- |
| {} |
| --- |
| # Vitals Interpreter Model (Fine-Tuned LLM) |
|
|
| ## Project Overview |
| This project implements a fine-tuned transformer model that interprets basic human vital signs and generates structured health guidance. |
|
|
| The model takes numerical vitals as input and produces a concise, human-readable output consisting of: |
| - Health status classification |
| - Suggested action/advice |
|
|
| --- |
|
|
| ## Objective |
| To build a lightweight, efficient AI system that: |
| - Understands structured vital inputs |
| - Classifies health condition into categories |
| - Generates consistent and controlled responses |
|
|
| --- |
|
|
| ## Model Details |
| - **Base Model:** t5-small |
| - **Architecture:** Encoder-Decoder Transformer |
| - **Fine-Tuning Type:** Supervised Fine-Tuning (SFT) |
| - **Framework:** Hugging Face Transformers |
|
|
| --- |
|
|
| ## Input Format |
| interpret vitals -> heart rate X, blood pressure Y/Z, temperature T |
|
|
|
|
| ### Example: |
| interpret vitals -> heart rate 125, blood pressure 150/95, temperature 100 |
|
|
|
|
| --- |
|
|
| ## Output Format |
| Status: <Normal | High | Low | Critical> | Advice: <short guidance> |
|
|
|
|
| ### Example Output: |
| Status: High | Advice: Monitor and consult doctor |
|
|
|
|
| --- |
|
|
| ## Dataset |
| - **Type:** Synthetic dataset |
| - **Size:** ~30–50 samples |
| - **Design Approach:** |
| - Based on medically accepted ranges of vital signs |
| - Balanced across categories: |
| - Normal |
| - High |
| - Low |
| - Critical |
|
|
| ### Why Synthetic Data? |
| Due to lack of publicly available labeled text datasets for this task, a controlled dataset was generated to: |
| - Ensure consistency in output format |
| - Improve learning efficiency |
| - Avoid noisy or unstructured data |
|
|
| --- |
|
|
| ## Training Configuration |
| - **Epochs:** 20–30 |
| - **Batch Size:** 2–4 |
| - **Learning Rate:** 5e-5 |
| - **Max Sequence Length:** 64 |
| - **Tokenizer:** AutoTokenizer (T5) |
|
|
| --- |
|
|
| ## Evaluation |
| ### Method: |
| - Manual testing with unseen inputs |
| - Verification of: |
| - Correct classification (Normal / High / Low / Critical) |
| - Proper output structure |
| - Relevance of advice |
|
|
| ### Sample Predictions: |
| | Input | Output | |
| |------|--------| |
| | HR: 125, BP: 150/95, Temp: 100 | Status: High \| Advice: Monitor and consult doctor | |
| | HR: 72, BP: 120/80, Temp: 98.6 | Status: Normal \| Advice: No action needed | |
| | HR: 140, BP: 170/110, Temp: 103 | Status: Critical \| Advice: Emergency care required | |
|
|
| --- |
|
|
| ## How to Use |
|
|
| ### Installation |
| ```python |
| from transformers import AutoTokenizer, AutoModelForSeq2SeqLM |
| |
| Author: |
| Archee Sinha |
| B.Tech CSE (AI) |
| ABES Institute of Technology |
