Machine learning models trained for high-stakes decisions face challenges when outcome data is missing. Historical decision-making influences observed outcomes, leading to biased predictions. A Bayesian model class with domain constraints improves risk estimation for both tested and untested patients. The prevalence and expertise constraints enhance parameter inference, as shown theoretically and on synthetic data. Applied to cancer risk prediction, the model accurately predicts diagnoses, aligns with public health policies, and identifies suboptimal test allocation.
A otro idioma
del contenido fuente
arxiv.org
Ideas clave extraídas de
by Sidhika Bala... a las arxiv.org 03-14-2024
https://arxiv.org/pdf/2312.03878.pdfConsultas más profundas