基于多模态自动化机器学习框架的住院患者跌伤风险预测模型的构建与验证

关键词：跌伤  住院患者  多模态自动化机器学习  风险预测模型  文本数据

Keywords:fall injuries  inpatients  multimodal automatic machine learning framework  risk prediction model  textual data

摘要：

      摘要 目的：基于多模态自动化机器学习（AutoML）框架，利用多模态临床大数据构建住院患者跌伤风险预测模型，进行内部和外部验证，并分析文本数据对跌伤风险预测的贡献。 方法：采用便利抽样法，选取2013年1月1日至2023年4月30日重庆市3家三级甲等综合医院的2，064例次住院患者跌伤事件为研究对象，共计369，048条数据。1家医院作为模型构建数据集，其他医院作为独立的验证数据集。采用AutoGluon-Tabular分析结构化数据，再融合多模态模型分析文本和结构化数据，通过permutation-shuffling分析跌伤相关因子特征，并基于多模态自动化机器学习框架构建预测模型，进行多中心外部验证。通过准确率、精确率、召回率、F1值和受试者工作特征曲线下面积（area under the receiver operating characteristic curve， AUC）检验模型的预测性能。 结果：以AutoGluon-Tabular分析结构化数据构建的模型准确率、精确率、召回率、F1值及AUC分别为0.676±0.024、0.663±0.066、0.676±0.024、0.622±0.026和0.7±0.049。通过融合AutoGluon-Tabular与多模态模型加入对文本数据的分析提高了模型性能。此外，外部验证表明，模型的准确率、精确率、召回率、F1值及AUC分别为0.679、0.642、0.679、0.654和0.664，与内部验证结果相似，模型具有较好的可重复性和外推性。同时，共发现68个住院患者跌伤风险相关因子，主要包括文本数据、自理能力等护理评估结果、骨疏康胶囊等药物、电解质等检验指标、胆囊结石等疾病诊断，以及Ⅰ、Ⅱ级手术等结构化数据。其中，现病史、既往史、个人史、主诉在特征重要性排名中靠前，表明文本数据对跌伤风险预测有重要意义。 结论：多模态AutoML框架能更好地处理临床多模态数据，采用此方法构建住院患者跌伤风险预测模型能较好地预测患者跌伤风险。

Abstract:

      Abstract Objective：To develop a risk prediction model for inpatient fall injury based on the multimodal automated machine learning （AutoML） framework with large-scale clinical data， conduct internal and external validation， and to analyze the contribution of text data to the prediction of fall injury risk. Method：Using convenience sampling, 2,064 cases of inpatient fall injury (369,048 data records) were collected from 3 tertiary comprehensive hospitals in Chongqing between January 1, 2013 and April 30， 2023. One hospital was served as the model construction dataset， while the others served as independent validation datasets. AutoGluon-Tabular was initially employed for structured data analysis. Subsequently， multimodal approaches were integrated to incorporate both text and structured information. The features of fall injury factors were analyzed through permutation shuffling and the prediction mode was constructed based on the multimodal AutoML for multi-center external validation. The predictive performance of the model was tested by accuracy， precision， recall， F1 value and area under the receiver operating characteristic curve （AUC）. Result：The accuracy， precision， recall， F1 score， AUC of the developed model for analyzing structured data were 0.676±0.024， 0.663±0.066， 0.676±0.024， 0.622±0.026， and 0.7±0.049， respectively. Integration of the AutoGluon-Tabular with multimodal techniques enhanced model performance by integrating textual data analysis. Moreover， external validation indicated that the model accuracy， precision， recall， and AUC were 0.679， 0.642， 0.679， 0.654 and 0.664， respectively， which were similar to the derivation cohort results， indicating promising reproducibility and extrapolation. A total of 68 factors associated with the risk of inpatient fall injuries were identified， including text data, nursing assessment results such as self-care ability, drugs such as gushukang capsules, experimental examination indicators such as electrolytes, diagnosis of diseases such as gallstones, grade I and II surgeries and other structured data. Overall， current medical history， past medical history， personal history， and chief complaint rank highly in terms of feature importance， indicating that text data are highly important for predicting the risk of inpatient fall-related injuries. Conclusion：The multimodal AutoML framework can better process the clinical multimodal data， and using this method to construct a risk prediction model can effectively predict inpatient fall injury risk.