Data and code availability: Machine Learning on systematically curated data reveals key determinants of magnetic hyperthermia performance doi-10-17044-scilifelab-29835419-0 https://doi.org/10.17044/SCILIFELAB.29835419 Swedish National Data Service Svensk nationell datatjänst Landing page Data and code availability: Machine Learning on systematically curated data reveals key determinants of magnetic hyperthermia performance doi-10-17044-scilifelab-29835419-0 https://doi.org/10.17044/SCILIFELAB.29835419 Vega, Edgar Teleki, Alexandra Swedish National Data Service Svensk nationell datatjänst Landing page The accurate prediction of the specific absorption rate (SAR) of superparamagnetic iron oxide nanoparticles (SPIONs) is critical for optimizing their performance in magnetic hyperthermia applications. This study presents the development of a predictive model for SAR using advanced machine learning techniques. A comprehensive dataset comprising 1,850 entries was compiled through the integration of 84 relevant scientific articles. The dataset listed 30 predictive features, including physical, chemical, and magnetic SPION properties, along with extrinsic experimental parameters commonly reported. Exploratory data analysis revealed complex nonlinear relationships among the predictive features. Twelve machine learning models were evaluated and refined using Bayesian hyperparameter optimization. The CatBoost algorithm emerged as the most effective model, achieving the lowest mean absolute error (20.92 W/g) and root mean squared error (39.41 W/g), along with a high coefficient of determination (R² = 0.98). Shapley Additive Explanation analysis identified the alternating magnetic field amplitude and frequency as the most influential factors, followed by SPION concentration and the surface area of the core nanoparticle. Conformal prediction analysis confirmed the model's reliability, providing a prediction interval of ±61.94 W/g. The model's generalization capability was validated using an independent dataset of SPIONs with varying sizes (from 7 nm to 30 nm) and dopants (Zn, Mn, Mg, and Co). The CatBoost model accurately predicted SAR values for small-sized nanoparticles (~7 nm), although predictions for medium (~15 nm) and large-sized (~30 nm) SPIONs exhibited greater variability. The study demonstrates that advanced machine learning models, such as CatBoost, can significantly contribute to the identification of nanoparticles with optimal properties for magnetic hyperthermia, thereby supporting their systematic and robust development for broader clinical use. Access to data through an external actor. Åtkomst till data via extern aktör.