Edit model card

Logo_Website

Thai Medical Large Language Model

Github : Github Evaluate
PaPer :

Llama-3.1-EIRAI-8B-instruct

Llama-3.1-EIRAI-8B-instruct: developed an 8-billion parameter model specifically tailored for Thai medical applications, with expertise in both Thai medical language and English medical terminology. The model has demonstrated its capabilities through key benchmarks such as MMLU, MedQA, PubMedQA, and MedMCQA, as well as Thai language assessments like ThaiExam, M3Exam, XNLI, and XCOPA. Additionally, we have created a Clinically Adapted Model Enhanced test using the Thai language to support clinical use in hospitals and to further improve the performance of Thai medical Retrieval-Augmented Generation (RAG).

Notice

While Eir AI Thai Medical LLM is designed to encode high-quality medical knowledge, it is not yet optimized for safe, practical use in real-world medical settings. The model is still in the research phase and should not be used for clinical decision-making without further validation, including randomized controlled trials. It is available for researchers to explore the potential of LLMs in medical contexts, but real-world deployment is not recommended in its current version.

Safety and Future Work

The current version of Eir AI Thai Medical LLM is under active development. We advise against using it for medical applications until further testing is completed. Our goal is to continue enhancing the model through rigorous testing and real-world evaluation, ensuring that it can be safely integrated into healthcare systems in the future.

Model Overview

  • Model Architecture: Meta-Llama-3.1-8B-Instruct
  • Version: 1.0
  • License(s): llama3.1

Evaluations

Medical Model Clinical KG Medical Genetics Anatomy Pro Medicine College Biology College Medicine MedQA PubMedQA MedMCQA Avg.
GPT-3.5 Turbo 1106 74.7 60.2 65.9 72.0 64.73 64.73 57.71 72.66 66.0 66.6
Thai LLMs
Eir AI-8B 75.1 80.0 69.6 76.8 77.1 66.5 64.5 79.0 58.6 71.9
Eir AI-8B + Prob 83.8 89.0 83.0 84.9 89.6 75.7 69.6 78.8 67.1 80.2
Typhoon-v1.5x-8B 75.9 79.0 63.7 70.6 77.1 63.6 59.7 74.4 58.0 69.1
OpenThaiGPT-beta-7B 37.4 38.0 4.5 32.7 36.1 32.4 32.4 62.0 31.8 34.1

Translation Performance Metrics

Model BLEU Score N-gram Precisions (%) BP Ratio
Typhoon-v1.5x-8B-Instruct 34.42 71.3/50.6/38.6/29.6 0.764 0.788
Meta Llama 3.1-8B Instruct 35.74 62.8/42.3/31.7/24.1 0.946 0.948
Eir AI-8B 61.10 76.1/64.6/56.6/50.1 1.000 1.006
Eir AI-8B-prob 47.91 74.0/58.0/48.2/40.6 0.890 0.896

Clinically Adapted Thai Medical Task Performance

Task GPT-3.5 Typhoon-v1.5x-8B-instruct GPT-4o Eir AI-8B
Named Entity Recognition 3.26 5.55 6.34 7.08
Temporal Information Extraction 3.83 5.46 6.15 7.05
Paraphrasing 2.36 4.68 6.35 7.06
Natural Language Generation 2.63 4.87 6.91 7.66
Keyword Extraction 2.60 5.15 7.01 7.35
Text Classification 2.92 6.21 5.36 6.75
Relation Extraction 3.29 5.94 4.37 6.92
Question Answering 3.70 4.92 6.11 6.82
Text Summarization 2.98 5.44 7.51 7.51
Abbreviation Expansion 3.99 5.96 6.24 7.82
Clinical Concept Normalization 2.67 5.63 5.82 6.55
Open-ended Question 3.32 5.55 6.77 7.27
Multiple-Choice Question 3.90 5.00 5.40 6.40
Coreference Resolution 3.48 4.55 4.88 6.43
Yes/No Question 2.71 5.86 4.86 7.38
Medical Translation 3.00 4.00 7.79 7.65
Medical Thai Extraction 2.81 7.16 8.62 8.16
Medical ICD Prediction 2.08 3.16 8.12 6.41
Average Score 3.05 5.33 6.38 7.11

Prompt Template

This model uses ChatML prompt template:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>

{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Example Clinical Adapted ICD 10 Prediction

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are responsible for accurately assigning ICD-10 codes and to diagnose and document medical records.
Your expertise ensures that healthcare providers are properly reimbursed and that patient care is well-documented.
In this scenario, you will be presented with a series of medical records and your task is to provide the correct ICD-10 code(s)
and ICD-9 CM in procedures based on the information provided.

<|eot_id|>
<|start_header_id|>user<|end_header_id|>

"Chief Complaint :5วันก่อนมารพ.มีไข้ ไอ มีเสมหะ มีน้ำมูก เหนื่อย ปวดเมื่อยตามตัว \r\n
Present illness : 5วันก่อนมารพ.มีไข้ ไอ มีเสมหะ มีน้ำมูก เหนื่อย ปวดเมื่อยตามตัว  มีน้ำมูก เลือดกำเดาจาากข้างขวา
ปฏิการกระทบกระแทก ไม่มีเจ็บคอ ไม่มีอาการอ่อนเพลีย มีอาการอ่อนเพลีย ไอมาก ไอตลอด มีอาการระคายคอ ปัสสาวะปกติ ไม่มีถ่ายเหลว
\r\n\r\nAllergy : |\r\n\r\nOther : no underlying disease\r\n\r\nPlan Treatment Day 1 of hospitalization : admit ward
\r\n\r\nReview of System { \r\n\r\n General :a thai adult female ,look sickness fatigue dry lip  moderate dehydration
\r\n Skin :no MP rash \r\n Eyes :not pale ,no icteric sclera \r\n Chest :secretion sound in both lung ,no crepitation , no wheezing \r
\n }
VitalSign First : {\n
BP : 117.0/63.0 mmHg\n
Pulse : 62.0 BPm\n
Temperature : 37.0 Celsius\n
Respiratory rate : 20.0\n
Weight : 50.000 kgs.\n
Height : 165.0 cm.\n
Painscore: N/A\n
O2SAT : 100\n}\n

Lab Results: \n

Electrolyte:Sodium (Na), Result : 143 mmol/L\r\n
Electrolyte:Potassium (K),Result : 3.8 mmol/L\r\n
Electrolyte:Chloride (Cl), Result : 108 mmol/L\r\n
Electrolyte:Bicarbonate (CO2),Result : 27.0 mmol/L\r\n
Creatinine (Serum):Creatinine, Result : 0.69 mg/dL\r\n
Creatinine (Serum):eGFR,Result : 100.41 ml/min/1.73 m^2\r\n
AST/SGOT:AST/SGOT, Result : 48 U/L\r\n
ALT/SGPT:ALT/SGPT, Result : 42 U/L\r\n
CBC:WBC Count,Result : 3.2 10^3/uL\r\n
CBC:RBC Count, Result : 3.57 10^6/uL\r\n
CBC:Hemoglobin (Hb), Result : 10.7 g/dL\r\n
CBC:Hematocrit (HCT),Result : 32.4 %\r\n
CBC:MCV, Result : 91 fL\r\n
CBC:MCH, Result : 30.0 pg\r\n
CBC:MCHC, Result : 33.0 g/dL\r\n
CBC:RDW-CV,Result : 12.9 %\r\n
CBC:Platelet Count, Result : 226 10^3/uL\r\n
CBC:Platelet Estimates, Result : Adequate \r\n
CBC:Neutrophils,Result : 41 %\r\n
CBC:Lymphocytes, Result : 46 %\r\n
CBC:Monocytes, Result : 9 %\r\n
CBC:Eosinophils, Result : 4 %\r\n
CBC:Basophils,Result : 0 %\r\n
COVID-19 Antigen (Rapid Test) For Admit:COVID-19 Antigen (Rapid Test), Result : Positive \r\n
Influenza A/B screening:Flu A,Result : Negative \r\n
Influenza A/B screening:Flu B, Result : Negative \n

X-ray Results: 15-05-2567 12:56 CHEST (PA, UPRIGHT)\r\n
\r\nHISTORY: Fever with productive cough and dyspnea for 5 days. \r\nCOMPARISON: None.\r\n
\r\n
FINDINGS:\r\n
TUBES/LINES: None.\r\nLUNGS: Normal lung volumes. Band-like opacity at right apical lung. Clear left lung.\r\n
PLEURA: No pneumothorax or effusion. Mild bilateral apical pleural thickening. \r\n
HEART AND MEDIASTINUM: Normal cardiac silhouette. Normal mediastinal contour.\r\n
BONY STRUCTURES: Mild scoliosis of the thoracic spine. \r\n\r\n
IMPRESSION:\r\n
- Band-like opacity at right apical lung, probably plate atelectasis or fibrosis\r\n- Mild bilateral apical pleural thickening"

<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>

Let's analyze the provided information and document the appropriate ICD-10 codes based on the patient's symptoms and diagnosis.

### Step-by-Step Analysis:

1. **Understand the Symptoms and Diagnosis**:
   - The patient presents with a range of symptoms including fever, cough with sputum, fatigue, nasal discharge (including blood),
and dyspnea. The lab tests returned a positive COVID-19 antigen test, while the influenza test was negative.

2. **Identify Key Terms**:
   - Primary key terms include 'COVID-19 positive', 'fever', 'productive cough', 'dyspnea', and pulmonary findings from the X-ray.

3. **Locate the Main Category**:
   - The main categories for the conditions noted are:
     - COVID-19: U07.1
     - Possible acute bronchitis: J20 (based on the productive cough)
     - Other symptoms such as dizziness: R42
     - Evidence of acute respiratory infection: J06.9

4. **Identify Specific Codes**:
   - Based on the diagnosis:
     - COVID-19 confirmed case: U07.1
     - Acute bronchitis: J20.9 (considering unspecified acute bronchitis)
     - Dizziness (if noted as significant and problematic): R42
     - Unspecified acute upper respiratory infection: J06.9

5. **Check for Additional Codes or Modifiers**:
   - Considering the details provided, the primary diagnosis is the confirmed COVID-19 infection, while bronchiolitis
and dizziness might be secondary diagnoses.

6. **Validate the Code Selection**:
   - Confirm that the selected codes reflect the patient's symptoms accurately, including acute bronchitis
and any other significant symptoms presented.

7. **Document the ICD-10 Code**:
   - The ICD-10 codes are documented for all noted conditions accordingly."

```json
Answer: {
    "Primary_Diagnosis": {
        "Code": "U071",
        "Description": "COVID-19, virus identified"
    },
    "Secondary_Diagnosis": [
        {
            "Code": "J20",
            "Description": "Acute bronchitis due to unspecified organism"
        },
        {
            "Code": "R42",
            "Description": "Dizziness and giddiness"
        },
        {
            "Code": "J069",
            "Description": "Acute upper respiratory infection, unspecified"
        }
    ],
    "ICD9_Procedure": []
}

Example Clinical Adapted Thai Medical Extraction

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

Task : Extract input the following patient information into output format Tone: the following medical text into
Thai in a fluent and elegant style.
Output Format.1.Age: \n2.Gender: \n3.Weight :\n4.Height : \n5.Chief Complaint: \n6.Symptoms and Signs: \n7.Medical History: \n
8.Current Medications:  \n9.Laboratory Results:  \n10.Imaging Findings:  \n11.Allergy: \n12.Drug Allergy:


<|eot_id|>
<|start_header_id|>user<|end_header_id|>
ผู้ป่วยของเราเป็นชายถนัดทั้งสองมือ อายุ 43 ปี มีประวัติการชักที่ไม่สามารถควบคุมได้มาเป็นเวลา 20 ปี ลักษณะการชักของเขามักจะรวมถึงการรู้สึกร้อนวูบวาบและอาการทางประสาทสัมผัสอื่น ๆ
ที่พัฒนาไปสู่การเคลื่อนไหวของกล้ามเนื้อที่มีจุดศูนย์กลางส่วนใหญ่ทางด้านขวา การตรวจหาสาเหตุของการชักรวมถึงการถ่ายภาพด้วยคลื่นแม่เหล็กไฟฟ้า (MRI) ซึ่งเผยให้เห็นเนื้องอกไขมันขนาดใหญ่ที่เส้นกลางสมอง
การพัฒนาไม่สมบูรณ์ของคอร์ปัสคาโลซัมบางส่วน และรอยโรคที่อยู่ใกล้เคียงในสมองส่วนหน้าซ้ายที่คาดว่าจะเป็นเนื้องอกกลีอาล (glial neoplasm) ตามลักษณะภาพถ่ายทางรังสี
รอยโรคในสมองส่วนหน้าซ้ายด้านหน้าและตรงกลางประกอบด้วยการกลายเป็นหินปูนแบบเป็นก้อนพร้อมการเพิ่มขึ้นของสัญญาณ FLAIR ที่กว้างขวางซึ่งเกี่ยวข้องกับไจรัสซิงกูเลตทั้งสองข้างและสมองส่วนหน้าซ้าย
(รูปที่ ).\n\nการจัดการทางการแพทย์ล้มเหลวในการควบคุมการชักของเขาและเขาถูกส่งต่อเพื่อหาทางเลือกในการรักษาด้วยการผ่าตัด รอยโรคที่เพิ่มขึ้นถูกสังเกตด้วยการถ่ายภาพเพิ่มเติมและขอบเขตของอาการบวมน้ำก็เพิ่มขึ้นด้วย
ความกังวลเกี่ยวกับการพัฒนาเนื้องอกกลีอาลที่เพิ่มขึ้นและการควบคุมการชักที่ไม่ดีทำให้มีการแนะนำให้ทำการผ่าตัด
การตัดสินใจถูกทำขึ้นเพื่อดำเนินการผ่าตัดนำทางด้วยระบบประสาทเพื่อตัดมวลที่เพิ่มขึ้นในสมองส่วนหน้าซ้ายและการตัดสมองส่วนหน้าบางส่วนโดยใช้การตรวจคลื่นไฟฟ้าสมองระหว่างการผ่าตัด
(intraoperative electroencephalogram - EEG), การทำแผนที่คอร์ติคอล (cortical mapping) และการตรวจวัดศักย์ไฟฟ้าที่เกิดจากการกระตุ้นประสาทรับความรู้สึก
 (somatosensory evoked potentials - SSEP)\n\nตัวอย่างที่ส่งไปตรวจทางพยาธิวิทยาแบบแช่แข็งในระหว่างการผ่าตัดพบว่ามีเส้นใยโรเซนธาล (Rosenthal fibers)
และการกลายเป็นหินปูนแบบเป็นจุดซึ่งคาดว่าจะเป็นเนื้องอกกลีอาล การประเมินทางพยาธิวิทยาแบบถาวรเผยให้เห็นเนื้องอกไขมัน (lipoma) และความผิดปกติของคอร์ติคอลแบบเฉพาะจุด
(focal cortical dysplasia) แบบ Palmini Type IA ในสมองที่อยู่ใกล้เคียง ความผิดปกติเล็กน้อยของโครงสร้างคอร์ติคอลและการเกิดกลีโอซิส (gliosis)
ในเนื้อสมองขาวที่เกี่ยวข้องสามารถเห็นได้ในคราบสีฮีมาโทซิลินและอีโอซิน (hematoxylin and eosin - H&E) และคราบสีโปรตีนกรดกลีอาลไฟบริลลารี (glial fibrillary acidic protein - GFAP)
และการย้อมสีโปรตีนเส้นประสาท (neurofilament protein - NFP) ในบริเวณที่เกิดกลีโอซิสไม่แสดงหลักฐานของเซลล์ประสาทที่ผิดรูป เซลล์ประสาทขนาดใหญ่ หรือเซลล์ลูกโป่ง (รูปที่ ).\n\n
การถ่ายภาพหลังการผ่าตัด (รูปที่ ) แสดงให้เห็นการตัดรอยโรคที่เพิ่มขึ้นใกล้เคียงทั้งหมดในสมองส่วนหน้าซ้ายและไจรัสซิงกูเลต การลดลงอย่างมากของอาการบวมน้ำที่เกิดจากหลอดเลือดถูกสังเกตในภาพ
FLAIR เนื้องอกไขมันที่เส้นกลางสมองไม่เปลี่ยนแปลง\n\nการฟื้นตัวหลังการตัดรวมถึงอาการเสริมในบริเวณมอเตอร์ซินโดรมระดับปานกลางรวมถึงอาการอ่อนแรงครึ่งซีก การพูดล่าช้า และการเปลี่ยนแปลงทางอารมณ์
การเปลี่ยนแปลงเหล่านี้ค่อยๆ หายไปและพบว่าผู้ป่วยมีความสมบูรณ์ทางระบบประสาทในการติดตามผลที่คลินิก การควบคุมการชักเกิดขึ้นทันทีหลังการตัด ยาต้านการชักถูกใช้อย่างต่อเนื่องและค่อยๆ
ถูกถอดออกในเดือนต่อมาโดยไม่มีการกลับมาเป็นซ้ำ ผู้ป่วยไม่มีการชักเป็นเวลา 3 ปีหลังการตัด การถ่ายภาพติดตามผล 18 เดือนพบการเปลี่ยนแปลง T2/FLAIR ที่คงที่พร้อมกับบริเวณเล็ก ๆ 


<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>
1. อายุ: 43 ปี\n
2. เพศ: ชาย\n
3. น้ำหนัก: ไม่ระบุ\n
4. ส่วนสูง: ไม่ระบุ\n
5. อาการหลัก: การชักที่ไม่สามารถควบคุมได้\n
6. อาการและสัญญาณ: รู้สึกร้อนวูบวาบ, อาการทางประสาทสัมผัส, การเคลื่อนไหวของกล้ามเนื้อที่มีจุดศูนย์กลางส่วนใหญ่ทางด้านขวา\n
7. ประวัติการรักษา: การจัดการทางการแพทย์ล้มเหลวในการควบคุมการชัก\n
8. ยาที่ใช้ในปัจจุบัน: ยาต้านการชัก\n9. ผลการตรวจทางห้องปฏิบัติการ: ไม่ระบุ\n
10. ผลการตรวจภาพ: MRI แสดงเนื้องอกไขมันขนาดใหญ่ที่เส้นกลางสมอง, การพัฒนาไม่สมบูรณ์ของคอร์ปัสคาโลซัมบางส่วน, รอยโรคที่อยู่ใกล้เคียงในสมองส่วนหน้าซ้ายที่คาดว่าจะเป็นเนื้องอกกลีอาล\n
11. อาการแพ้: ไม่ระบุ\n
12. อาการแพ้ยา: ไม่ระบุ\n
การผ่าตัดนำทางด้วยระบบประสาทเพื่อตัดมวลที่เพิ่มขึ้นในสมองส่วนหน้าซ้ายและการตัดสมองส่วนหน้าบางส่วนได้ถูกดำเนินการ
และผลการตรวจทางพยาธิวิทยาเผยให้เห็นว่าเป็นเนื้องอกไขมันและความผิดปกติของคอร์ติคอลแบบเฉพาะจุด (focal cortical dysplasia) แบบ Palmini Type IA
ในสมองที่อยู่ใกล้เคียง การฟื้นตัวหลังการผ่าตัดรวมถึงอาการเสริมในบริเวณมอเตอร์ซินโดรมระดับปานกลาง แต่การควบคุมการชักเกิดขึ้นทันทีหลังการผ่าตัดและผู้ป่วยไม่มีการชักเป็นเวลา
3 ปีหลังการผ่าตัด การถ่ายภาพติดตามผล 18 เดือนพบการเปลี่ยนแปลง T2/FLAIR ที่คงที่พร้อมกับบริเวณเล็ก ๆ ที่เพิ่มขึ้นของการเพิ่มขึ้นใกล้เคียงที่ไม่เปลี่ยนแปลง.

How to use


from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

model_id = "EIRTHAIMED/Llama-3.1-EIRAI-8B"

nf4_config = BitsAndBytesConfig(
   load_in_4bit=True,
   bnb_4bit_quant_type="nf4",
   bnb_4bit_use_double_quant=True,
   bnb_4bit_compute_dtype=torch.bfloat16
)

# Load the base model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    # quantization_config=nf4_config, # uncomment this line for 4 bit loading
    device_map="auto",
    attn_implementation="flash_attention_2"
)

messages = [
    {"role": "system", "content": "You are an expert medical assistant named EIR , developed by EIR Thai Medical LLM. You are to be a helpful, respectful, and honest assistant."},
    {"role": "user", "content": "การใช้ clinical tracer มีบทบาทอย่างไรในการพัฒนาคุณภาพการดูแลผู้ป่วย?"}
]

input = tokenizer.apply_chat_template(
    messages,
    tokenize = True,
    add_generation_prompt = True, # Must add for generation
    return_tensors = "pt",
).to("cuda")


from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer, skip_prompt = True)
_ = model.generate(input, streamer = text_streamer, max_new_tokens = 1500, do_sample=True, temperature=0.01, top_k=100, top_p=0.95) 

@article{EirAI,
    title={Eir: Thai Medical Large Language Models}, 
    author={Yutthakorn Thiprak and Rungtam Ngodngamthaweesuk and Songtam Ngodngamtaweesuk, MD},
    year={2024},
    journal={arXiv preprint arXiv:2409.08523},
    url={https://arxiv.org/abs/2409.08523}
}

Thank you very much Asst.Prof.Dr. Ekapol Chuangsuwanich and Praj Bhargava @Meta Research Engineer, for your valuable endorsement of our preprint paper on arXiv.

Thank you Draft Reviewer Report Kullawat Chaowanawatee and Dr. Jakapan Suaboot from Prince of Songkla University, Phuket Campus
Draft Industry Reviewer Report Mr. Piyawat Maneenual ,Assistant IT Manager ,Thonburi Rajyindee Hospital

Downloads last month
204
Safetensors
Model size
8.03B params
Tensor type
BF16
·
Inference Examples
Inference API (serverless) is not available, repository is disabled.

Model tree for EIRTHAIMED/Llama-3.1-EIRAI-8B

Finetuned
this model
Quantizations
2 models