基于公众判断的癌症患者安宁转诊智能分类预测模型的构建

Development of an artificial intelligence model for predicting palliative care referral of cancer patients based on public judgment

目的构建基于公众判断的癌症患者安宁转诊智能分类预测模型,为癌症患者的安宁转诊决策提供参考。方法使用质性研究与量性研究相结合的方法确定基于公众判断的安宁转诊指标;回顾性收集354例晚期癌症患者的临床及随访资料,采用随机森林分类算法构建癌症患者安宁转诊预测模型。采用前瞻性研究方法收集75例晚期癌症患者相关资料用于外部验证,并根据结果建立网页计算器。结果基于公众判断的癌症患者安宁转诊预测模型包括疼痛、营养不良、呼吸困难、意识水平下降、自理能力受限、活动受限、腹水、水肿、抑郁、排便排尿障碍、吞咽困难、难以处理的伤口12个指标,转诊时间宜预计生存期<3个月。预测模型的ROC曲线下面积为0.876,Brier分值为0.128;模型外部验证ROC曲线下面积为0.831,Brier分值为0.095。结论癌症患者安宁转诊分类预测模型具有较好的预测效能,可为癌症患者安宁转诊的快速评估提供参考。

Objective To construct an artificial intelligence(AI)model for predicting palliative care referral of cancer patients based on public judgment,and to provide reference for referring cancer patients to a palliative care service.Methods A mixed-method(qualitative and quantitative)approach was used to identify indicators for palliative care referral of cancer patients based on public judgment.Clinical and follow-up data from 354 advanced cancer patients were retrospectively collected,and Random Forest algorithm was applied to develop the prediction model.Additionally,75 terminal cancer patients were prospectively studied for external validation of the prediction model,and a web-based calculator was developed based on the model results.Results The prediction model for palliative care referral of cancer patients included 12 indicators:pain,malnutrition,respiratory distress,decreased consciousness level,limited self-care ability,restricted mobility,ascites,edema,depression,defecation and urination difficulties,dysphagia,and difficult-to-heal wounds.The optimal timing for referral should be with an estimated life expectancy of less than three months.The area under the receiver operating characteristic curve(AUC)for the prediction model was 0.876,and the Brier score was 0.128.The AUC in the external validation sample was 0.831,and the Brier score was 0.095.Conclusion The public judgment-based prediction model for palliative care referral of cancer patients demonstrates satisfactory predictive performance and can serve as a rapid screening tool for palliative care referral.