Prediction of Tuberculosis Using an Automated Machine Learning Platform for Models Trained on Synthetic Data

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Biomarker Analysis

Background:

• High-quality medical data is essential for developing machine learning (ML) algorithms in healthcare.
• Privacy and security concerns restrict access to real-world medical data, hindering ML development.
• Synthetic data offers a potential solution to overcome data access limitations.

Purpose of the Study:

• To evaluate the utility of synthetic data for training ML algorithms.
• To assess ML model performance for tuberculosis detection using inflammatory biomarker profiles.
• To compare the effectiveness of models trained on real versus synthetic datasets.

Main Methods:

• Generated synthetic datasets (B, C, D) from a retrospective real dataset (A) of 278 patients.
• Trained ML models on both real (Dataset A) and synthetic datasets.
• Validated model performance using accuracy, sensitivity, and specificity metrics.

Main Results:

• ML models trained on real data achieved 90% accuracy, 89% sensitivity, and 100% specificity.
• Models trained on the optimal synthetic dataset (B) showed 91% accuracy, 93% sensitivity, and 77% specificity.
• Performance varied across synthetic datasets, with datasets C and D showing lower accuracies (71% and 54%).

Conclusions:

• Synthetic data can serve as a viable alternative for training ML algorithms in healthcare.
• This approach may expedite the development and implementation of ML tools for disease detection.
• Further research is warranted to optimize synthetic data generation for robust ML model training.