Groundwater health probability risk prediction through oral intake using advanced optimization methods.

Area of Science:

• Environmental Health
• Data Science
• Computational Biology

Background:

• Groundwater (GW) is a primary source for drinking and agriculture in many regions.
• Assessing cancer risk from oral GW intake is critical for public health.
• Generative AI offers potential for data augmentation but requires validation.

Purpose of the Study:

• To evaluate the reliability of generative AI in producing scientific data for cancer risk prediction.
• To compare the predictive performance of various machine learning models using real versus synthetic GW data.
• To determine the optimal models for predicting cancer risk from GW ingestion.

Main Methods:

• Integrated real experimental data with generative AI-driven synthetic data.
• Evaluated standalone models: Artificial Neural Networks (ANN), Gaussian Process Regression (GPR), Support Vector Machines (SVM), and Boosted Trees (BT).
• Utilized Bayesian Optimization (BO) to enhance model performance and compared results on real and synthetic datasets.

Main Results:

• On real data, ANN and ANN-BO models demonstrated superior performance with low errors (e.g., ANN-BO testing MAE=0.0902).
• Models trained and tested on synthetic data exhibited significantly higher errors (e.g., ANN-BO testing MAE=15.718), indicating poor generalization.
• Bayesian Optimization improved model performance, especially on real data, but could not overcome synthetic data limitations.

Conclusions:

• Real-world experimental data is indispensable for accurate cancer risk prediction from GW ingestion.
• Generative AI-produced synthetic data, while useful for some applications, is currently insufficient for reliable health risk assessments.
• Data quality and validation using real-world evidence are paramount for minimizing errors and ensuring the credibility of predictive models in environmental health.