Logotyp: till Uppsala universitets webbplats

Magnetic hyperthermia is a supplementary therapy employed to improve cancer treatment, where Superparamagnetic Iron Oxide Nanoparticles (SPIONs) play a crucial role in many applications. However, the development of SPIONs is a time-consuming activity and involves considerable resources. This difficulty is primarily attributed to the complex and not-fully understood relationships between the SPIONs heating efficiency and their physicochemical and magnetic properties. This study aims to address this problem by constructing a relevant dataset based on previous literature and developing a robust predictive model capable of capturing these intricate relationships. To achieve this goal, natural language processing techniques were applied to build a relevant dataset, and ten machine learning algorithms were explored to predict the Specific Absorption Rate (SAR) of SPIONs based on their physicochemical and magnetic properties.

Among the models tested, the Support Vector Regressor (SVR) model reported the most promising results, demonstrating superior predictive performance after hyperparameter tuning. In the final evaluation, the SVR model showed robust generalization with a Root Mean Square Error (RMSE) value of 51.2555 on the test data and 43,2160 on the training data. The employment of  Box-Cox transformation proved to be the most effective strategy for enhancing the SVR model performance.  The external magnetic field applied was found to be the most significant predictor in the model, followed by saturation magnetization and the SPION core diameter. Interestingly, non-linear models outperformed linear models, suggesting that the relationships between the predictor and target features are inherently complex and mostly non-linear. This result highlights the need to implement non-linear machine-learning models to capture the relationships of the features and accurately predict the SAR value.

This study offers a promising approach for optimizing the development of  SPIONs by focusing on influential magnetic and physicochemical properties. Future research efforts should further investigate non-linear-based machine learning models, contributing to improving hyperthermia treatment.