preprint · research square · 2026

Reference-Class Failure in Mortality Prediction: A Synthetic Study of Irreducible Uncertainty in End-of-Life Prognostic Models

abstract

Defines reference-class failure in mortality prediction: an end-of-life case can sit outside the development population’s transportable conditional mechanism while the model still returns a sharp individual risk. In a synthetic concept-shift study, logistic regression, random forest, gradient boosting, and a neural network remain confident on novel terminal cases, but Brier error plateaus near 0.215, far above the novel regime’s 0.092 aleatoric floor. The paper proposes a diagnostic triad: absent reference class, sharp output, and data-insensitive error.

@misc{nguyen2026referenceclass,
  author = {Nguyen, Michel},
  title = {Reference-Class Failure in Mortality Prediction: A Synthetic Study of Irreducible Uncertainty in End-of-Life Prognostic Models},
  year = {2026},
  howpublished = {Research Square},
  doi = {10.21203/rs.3.rs-10058406/v1},
  url = {https://doi.org/10.21203/rs.3.rs-10058406/v1}
}