机器学习在造口患者管理中应用的范围综述

Machine learning applications in the management of stoma patients: a scoping review

摘要:

背景与目的 随着造口患者数量不断上升，其长期管理需求呈现复杂化趋势。机器学习（ML）因擅长处理多维度临床数据，被逐渐应用于造口患者评估与护理。本研究通过范围综述系统梳理ML在造口领域的应用现状，明确研究热点、性能表现及存在的关键问题，为后续研究和临床转化提供依据。方法 采用澳大利亚乔安娜布里格斯研究所范围综述指南，系统检索中英文数据库建库至2025年4月10日的相关文献。由2名研究者独立筛选、提取及归纳研究特征、ML类型、应用领域、模型预测因子与性能，并使用PROBAST+AI工具评价方法学质量和偏倚风险。结果 共纳入15篇文献，ML在造口领域的应用包括术后并发症预测、疾病预后评估、健康教育、护理知识库构建、卫生经济学分析及身体成分评估等。10项研究共构建40个预测模型，覆盖6类预测因子（人口学、生理/身体状况、手术因素、疾病分期与治疗、造口相关情况、心理社会因素），多数模型在训练集中表现出较高AUC（多数>0.80）。然而，PROBAST+AI评价显示多数模型存在高偏倚风险，尤其在样本量、数据完整性和外部验证不足方面。LLM辅助健康教育显示出提高信息可获得性的潜力，但阅读难度较高且缺乏个性化。经济学研究提示ML有助于优化预防性造口策略并节约成本。结论 ML已在造口患者管理的多环节展现应用价值，尤其在风险预测与临床决策支持方面具有明显优势。但现阶段研究整体方法学质量有待提高，外部验证、模型可解释性与临床可实施性不足仍限制其推广。未来需开展大样本、多中心前瞻性研究，拓展至泌尿造口等更广泛人群，并推动个性化健康教育与智能化护理工具的临床落地。

Abstract:

Background and Aims The growing population of patients living with a stoma has led to increasingly complex clinical and long-term management needs. Machine learning (ML), with its strong capability for processing multidimensional data, has been progressively applied to stoma care. This scoping review aims to map current applications of ML in stoma patients, summarize research trends and model performance, and identify existing gaps to support future methodological and clinical development.Methods Guided by the Joanna Briggs Institute scoping review methodology, eight major Chinese and international databases were searched from inception to April 10, 2025. Two reviewers independently screened the literature, extracted data on study characteristics, ML algorithms, application domains, predictors, and model performance. Methodological quality and risk of bias were assessed using the updated PROBAST+AI tool.Results Fifteen studies were included. ML has been applied in multiple domains, including postoperative complication prediction, prognosis assessment, patient education, nursing knowledge-base construction, health economic analyses, and body composition assessment. Ten studies developed a total of 40 prediction models covering six categories of predictors: demographic characteristics, physiological/physical status, surgical factors, disease staging and treatment, stoma-related variables, and psychosocial factors. Most models demonstrated good discriminative ability (AUC>0.80). However, PROBAST+AI revealed generally high risk of bias, mainly due to retrospective designs, incomplete data reporting, and insufficient external validation. Large language model-based educational tools improved information accessibility but showed limited personalization and high reading difficulty. Economic studies suggested that ML-assisted decision-making may offer cost-effective support for protective stoma strategies.Conclusion ML shows promising value across multiple stages of stoma management, especially in risk prediction and decision support. Nevertheless, methodological limitations-particularly high bias risk, limited external validation, and inadequate clinical integration-hamper its broader application. Future research should prioritize multicenter prospective studies, expand to diverse stoma populations including urostomy, enhance model interpretability, and promote clinically deployable intelligent care tools.