In February 2025, the journal Sci published a study by Hungarian researchers – including Dr. Tamás Joó, Deputy President of MEMT – carried out within the framework of the Data-Driven Health Division of the National Laboratory for Health Security. The research examined the reliability of machine learning models used for survival analysis in cases where data are incomplete or “censored,” meaning that the exact survival time of all patients is not known.
The authors developed a new testing method that systematically measures how model performance changes as the proportion of missing data increases. Based on the evaluation of five models, they found that nonlinear approaches – such as mixture density networks – are particularly sensitive to data quality deterioration.
This method offers a new benchmark for assessing survival models and provides practical guidance for researchers and clinicians in selecting and developing more reliable predictive models. The study highlights that robust strategies for handling censored data are essential in clinical applications.
The full article is available at the following link: https://www.mdpi.com/2413-4155/7/1/18