选择性共形预测在乳腺癌高风险人群风险评估中的方法研究

Selective Conformal Prediction for Risk Assessment in High-Risk Breast Cancer Subgroups

摘要: 在医学风险预测中，为个体化预测结果提供可靠的不确定性量化，是评估模型临床实用性的关键环节。本文聚焦于选择性共形预测方法在乳腺癌高风险人群风险评估中的应用，旨在通过构建覆盖率可控的预测区间，实现对风险模型预测性能的稳健评估，并提升其在高风险人群中的可解释性。我们首先系统回顾了分裂共形预测的基本原理，进一步引入选择性共形预测方法，在保证理论覆盖率的同时，实现高风险个体的有效识别与重点覆盖，并针对该人群给出最大置信度与预测可信度等单指标不确定性度量，为临床干预提供参考信息。结合基于临床指标的风险预测模型，本文利用威斯康星州乳腺癌数据开展实证分析。结果表明，所提出的方法在高风险个体的预测准确率与不确定性控制方面均优于分裂共形预测，具有良好的实际应用潜力。本研究为医学诊断场景下的不确定性建模提供了新的统计工具，并为高风险人群的精准防控提供了方法上的支持。

Abstract: In biomedical research, reliable quantification of predictive uncertainty is a key criterion for evaluating the clinical utility of risk prediction models. This study focuses on the application of selective conformal prediction to risk assessment in high-risk breast cancer populations. The objective is to construct prediction sets with controllable coverage, enabling robust evaluation of model performance and enhancing interpretability for high-risk subgroups. We first review the split conformal prediction method and then introduce a selective conformal prediction framework that achieves both theoretical coverage guarantees and effective identification with prioritized coverage of high-risk individuals. For these individuals, we provide uncertainty measures such as maximum confidence and credibility scores to inform targeted clinical interventions. Using a risk prediction model incorporating clinical indicators, we conduct real data analysis based on the Wisconsin Breast Cancer dataset. Results demonstrate that the proposed method outperforms standard split conformal prediction in terms of predictive accuracy and uncertainty control for high-risk cases, showing strong potential for practical applications. This work offers a novel statistical tool for uncertainty modeling in medical diagnostics and provides methodological support for precision prevention and control in high-risk populations.