Synthetic Healthcare Datasets for SOAP Note Generation

This dataset collection provides synthetic healthcare data and associated training files developed for training and evaluating four specialized TinyLLaMA models for automated SOAP (Subjective, Objective, Assessment, Plan) note generation in rural / remote healthcare settings. The datasets were created as part of the project "Empowering Remote Healthcare with On-Premises Solar-Powered AI Units: Design and Implementation," aimed at developing affordable, sustainable AI solutions for remote regions with limited infrastructure. The collection includes synthetic patient records, transcriptions, preprocessed training files for SOAP sections, and a detailed data analysis report. Dataset Components: transcriptions.jsonl (1,031 records):This file contains synthetic transcriptions of patient-provider interactions, generated using GPT-4o (model version: gpt-4o-2024-08-06). Each record includes a custom_id (e.g., row0000) and a transcription field capturing a dialogue between a healthcare provider and a patient. The transcriptions simulate a variety of clinical scenarios, covering symptoms, medical history, lifestyle factors, physical examinations, and care plans. The dataset has been cleaned to remove records with invalid JSON formatting, ensuring all transcriptions are usable for downstream applications. soaps.jsonl (5,160 records):This file contains synthetic patient records paired with pre-generated SOAP note sections. Each record includes fields such as row_id, gender, age, smoker, drinker, blood_pressure, chief_complaint, subjective, objective, assessment, and plan. The data was synthetically generated and augmented using GPT-4o to simulate realistic patient profiles and corresponding clinical notes, with a focus on diversity in demographics and health conditions. Training JSONL Files for SOAP Sections:These files were derived from an augmented dataset (soap_augmented.jsonl) by splitting paraphrased SOAP sections into four subsets, each corresponding to a SOAP section used for fine-tuning individual TinyLLaMA models: subjective_for_training.jsonl (5,160 records): Paraphrased Subjective sections. objective_for_training.jsonl (5,160 records): Paraphrased Objective sections. assessment_for_training.jsonl (5,160 records): Paraphrased Assessment sections. plan_for_training.jsonl (5,160 records): Paraphrased Plan sections.Each file contains format (e.g., simple, bullets, narrative) and text fields, representing different paraphrasing styles to enhance model robustness. Data Analysis for Solar-Powered AI Healthcare Units.pdf:This PDF report provides a comprehensive analysis of the synthetic datasets, including demographic distributions (gender, age, smoker, drinker status), chief complaint frequencies, SOAP note structure, preprocessing steps for fine-tuning, bias analysis, and uncertainties. It also outlines the rationale for fine-tuning four TinyLLaMA models, model selection (TinyLLaMA 1.1B), fine-tuning process (3 epochs, batch size 8, learning rate 2e-5, AdamW optimizer), and ethical considerations. The report emphasizes the need for real-world validation and invites public critique to refine the approach. Purpose:These datasets and training files were developed to support the training of four TinyLLaMA models (aschbat/subjective-tinyllama, aschbat/objective-tinyllama, aschbat/assessment-tinyllama, aschbat/plan-tinyllama) for generating SOAP note sections from transcriptions. They are intended for use in research and development of AI-driven healthcare solutions, particularly for low-resource settings where automated medical documentation can improve access to care. The data analysis report provides insights into the dataset’s structure, potential biases, and fine-tuning methodology, serving as a resource for researchers and practitioners. Usage: Format: All data files are in JSONL (JSON Lines) format, with one JSON object per line. The analysis report is a PDF. Fields: generated_transcriptions_cleaned.jsonl: custom_id (string, e.g., row0000), transcription (string, patient-provider dialogue). synthetic_patients_with_soap.jsonl: row_id (string), gender (string), age (integer), smoker (string), drinker (string), blood_pressure (string), chief_complaint (string), subjective (string), objective (string), assessment (string), plan (string). Training JSONL files (subjective.jsonl, objective.jsonl, assessment.jsonl, plan.jsonl): format (string, paraphrasing style), text (string, paraphrased SOAP section content). Applications: The datasets can be used for training natural language processing (NLP) models for tasks such as text generation, information extraction, and clinical note automation. They are particularly suited for research in medical NLP, healthcare AI, and rural healthcare solutions. Preprocessing: Users may need to parse the JSONL format (e.g., using Python’s pandas.read_json(lines=True) or equivalent). For generated_transcriptions_cleaned.jsonl, the transcription field can be directly used as input for NLP models. The training JSONL files are preprocessed for fine-tuning on specific SOAP sections. This dataset was used to train four TinyLLaMA models for SOAP note generation, available on Hugging Face: Subjective: https://huggingface.co/aschbat/subjective-tinyllama Objective: https://huggingface.co/aschbat/objective-tinyllam Assessment: https://huggingface.co/aschbat/assessment-tinyllama Plan: https://huggingface.co/aschbat/plan-tinyllama These models are designed for rural / remote healthcare settings using solar-powered AI units. We welcome feedback and collaboration to refine their performance in real-world applications. Citation Requirement:When using this dataset, please cite: Schmidt Batista, Adans, (2025). Synthetic Healthcare Datasets for SOAP Note Generation: Transcriptions, Patient Records, and Training Data for Solar-Powered AI Healthcare Units. Zenodo. 10.5281/zenodo.15399846. License:This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, which permits users to share and adapt the data for any purpose, provided appropriate credit is given to the author (Schmidt Batista, Adans), a link to the license is provided, and any changes are indicated. Source and Generation:The data was synthetically generated using GPT-4o (OpenAI) as part of a project to develop AI-driven healthcare solutions for rural/remote settings. The transcriptions simulate realistic patient-provider interactions, while the patient records and SOAP sections were augmented to include diverse clinical scenarios and paraphrasing styles. The transcriptions.jsonl dataset was cleaned to ensure valid JSON formatting, with one problematic row removed from an original 1,032 records. Analysis and Findings:The accompanying data analysis report (Data Analysis for Solar-Powered AI Healthcare Units.pdf) provides insights into the dataset: Demographic Distribution: Gender (Female: 34.88%, Male: 35.27%, Non: 29.84%), Smoker (Yes: 49.42%, No: 50.58%), Drinker (Yes: 50.48%, No: 49.52%). Chief Complaints: Top complaints include bacterial infections, depression, skin rashes, and breast pain, reflecting diverse health concerns. Bias and Uncertainties: The dataset shows balanced demographic distributions but may lack the nuanced variability of real clinical encounters due to synthetic generation. Potential biases in language patterns and the risk of overfitting to synthetic data necessitate real-world validation. Fine-Tuning: The four TinyLLaMA models were fine-tuned on specific SOAP sections using the Hugging Face Transformers library, with each model optimized for deployment on low-power devices like the Raspberry Pi 5 (requiring ~2GB memory in 4-bit quantization). Limitations: The data is synthetic and does not contain real patient information. While designed to be realistic, it may not fully capture the nuances of actual clinical interactions. The transcriptions.jsonl dataset has 1,031 records after cleaning (originally 1,032, with one row removed due to JSON parsing errors). Users should validate model outputs for clinical accuracy if applied in real-world settings, as synthetic data may introduce artifacts like standardized language. Keywords:healthcare, synthetic data, SOAP notes, machine learning, NLP, rural healthcare, transcriptions, patient records, TinyLLaMA, AI in medicine, solar-powered AI, data analysis, fine-tuning Funding and Acknowledgments:This dataset was created as part of the project "Empowering Remote Healthcare with On-Premises Solar-Powered AI Units: Design and Implementation." Public feedback is encouraged to refine the approach for real-world deployment.