A joint research team composed of Professor Kwangsoo Kim of Jeonbuk National University (Department of Statistics; Vice‑Chair of the Korean Association for Artificial Intelligence) and Professor Sangwook Kang of Yonsei University (Department of Statistical Data Science) has developed a new calibration algorithm that improves the probabilistic calibration of survival analysis models: KSP (Kolmogorov-Smirnov metric-based Post-Hoc Calibration).

The study was accepted for presentation at NeurIPS 2025, a leading conference in artificial intelligence to be held in December 2025 in San Diego and Mexico City, USA.

In high-risk domains such as medicine and biotechnology, the calibration of predicted probabilities is as important as the predicted values. Along with discrimination—identifying which individuals are at greater risk—accurately predicting true survival probabilities determines the reliability of survival analysis applications.

Although the introduction of deep neural networks has improved discrimination, the mismatch between true risk probabilities and predicted probabilities has tended to increase. The research team therefore focused on the Kolmogorov-Smirnov metric (KS metric) and proposed an improved measure called KS-cal together with the KSP method.

KSP is a simple and efficient post-hoc calibration framework that transforms the predicted cumulative distribution function using a monotonically increasing link function, such as the logit, and minimizes KS-cal.

KSP achieved higher probabilistic calibration than existing methods with almost no sacrifice of already secured discrimination. Unlike other methods, KSP requires no additional steps such as partitioning intervals or sampling, so inference time does not increase substantially with sample size.

Across 60 experimental settings—combining six models (including DeepSurv and MTLR) with 10 benchmark datasets such as WHAS and METABRIC—KSP showed the best performance in about 70% of cases. In the remaining cases, it delivered comparable performance, demonstrating robustness.

Moreover, KSP offers ease of use and efficiency for survival analysis researchers, and is expected to increase the accuracy and trustworthiness of AI-based risk identification.

Professor Kwangsoo Kim of the Department of Statistics at JBNU said, 'It has been confirmed that around 30,000 papers were submitted to NeurIPS 2025 by AI researchers worldwide. This will likely be the largest conference in history. I am pleased that our paper was accepted to such a conference and that research we have worked on for two to three years is bearing fruit,' adding, 'We look forward to continued collaboration with researchers at home and abroad and ongoing support from various institutions.'

This research was supported by the National Research Foundation of Korea with funding from the Ministry of Science and ICT, and was carried out with support from the National Research Foundation's G-LAMP (LAMP) project funded by the Ministry of Education.