背景
慢性疼痛是影响生活质量的常见问题。通过血液生物标志物和临床变量，可以构建预测模型辅助临床决策。
技术架构
采用经典的机器学习流程：
- 数据预处理: 缺失值处理、特征工程
- 模型选择: 多种分类器比较（LR, RF, XGBoost）
- 评估指标: AUROC、敏感性、特异性、校准曲线
- 不确定性量化: Bootstrap 置信区间
关键特性
模型校准
使用校准曲线评估预测概率的可靠性，确保模型输出可用于临床决策。
临床评价指标
除了统计指标，还计算了临床需要的敏感性/特异性阈值下的阳性和阴性预测值。
解释性
通过 SHAP 值分析特征贡献，提供临床可解释性。
成果
- 构建了可解释的疼痛分类模型
- 提供了完整的模型比较和评估流程
- 模型校准良好，预测概率可靠
Background
This research project explores whether blood biomarkers and clinical variables can support chronic-pain classification in a small-sample setting.
Method
The workflow covers missing-data handling, feature preparation, comparison of several classical classifiers, bootstrap confidence intervals, and model calibration. Evaluation includes AUROC, sensitivity, specificity, and predictive values at clinically relevant thresholds.
Interpretation
SHAP analysis is used to inspect feature contributions. Calibration curves assess whether predicted probabilities remain meaningful beyond ranking performance.
Status
This is a completed biomedical machine-learning study and an earlier part of the portfolio. Current work has moved toward AI infrastructure, execution reliability, and developer tools.