AIPFR Prediction app

Introduction: Idiopathic pulmonary fibrosis is clinical entity defined by specific radiological and/or histological criteria. It is clear that IPF disease behaviour may be vastly different in each individual. Prediction of an individual’s disease trajectory is crucial to their management, but a mechanism by which to do this is currently lacking.Aims: In the current study, using data from the Australian IPF registry (AIPFR), we aimed to identify a panel of the most predictive factors useful for disease progression. Methods: Specifically, we developed the outcome of interest as positive if either a 10% drop in FVC, a 15% drop in DLCO, transplant or death was observed, and negative otherwise. To define the panel’s utility, we designed one set of analyses specific to a patients first presentation, and a second specific to a follow up presentation, where change in FVC and DLCO in the last 6 months were known. Using a suite of machine learning algorithms, we assessed approximately250 different demographic and clinical risk factors measured by the AIPFR to define the risk of progression in the next 12 months. Final predictive capability was measured via a combination of logistic regression and receiver operating characteristic (ROC) analyses. Results: The optimal combination of risk factors to predict progression upon first presentation included exposure to prednisone, patient’s percent predicted DLCO and FVC, clubbing, WHO functional class and two individual components from the SGRQ; medication side effects and feeling breathless walking at home. This combination resulted in an AUC of 0.75 (95%CI:0.7–0.79), with sensitivity, specificity and accuracy for the model at 61%, 79% and 70% respectively. To predict progression at follow-up, features included 6-month change in raw DLCO & FVC, percent predicted DLCO, cough severity on visual analogue scale, pulmonary hypertension, and two components of the SGRQ: medication side effects and a feeling of public discomfort. Predictive values increased, with the AUC reaching 0.813 (95%CI:0.78–0.85), with sensitivity, specificity and accuracy for the model at 75%, 72% and 73% respectively.Innovative contribution to policy, practice and/or research: Using a data driven approach to identify markers of progression this work has defined a panel of features that can be routinely followed up by practitioners to identify probability of progression in the next 12 months.Lineage: The app was written using R statistical software, and implemented using the R Shiny package. Data was obtained through Australian IPF registry.引言：特发性肺纤维化（Idiopathic Pulmonary Fibrosis, IPF）是一类以特定影像学和/或组织病理学标准定义的临床病症。现已明确，IPF的疾病表型在不同个体间存在显著差异。对个体患者的疾病轨迹进行预测对其临床管理至关重要，但目前尚缺乏可靠的实现方法。 研究目的：本研究依托澳大利亚IPF注册研究（Australian IPF Registry, AIPFR）的数据，旨在筛选出可用于预测疾病进展的最优预测因子组合。 研究方法：具体而言，本研究将研究结局指标定义为阳性：即出现用力肺活量（Forced Vital Capacity, FVC）下降10%、肺一氧化碳弥散量（Diffusing Capacity of the Lung for Carbon Monoxide, DLCO）下降15%、肺移植或死亡任一情况；反之则为阴性。为评估该预测因子组合的临床效用，我们设计了两类分析场景：其一针对患者的首次就诊数据，其二针对已知近6个月FVC与DLCO变化情况的随访就诊数据。我们采用一系列机器学习算法，对AIPFR收录的约250项人口学与临床危险因素进行评估，以明确患者未来12个月的疾病进展风险。最终的预测性能通过逻辑回归与受试者工作特征（Receiver Operating Characteristic, ROC）分析相结合的方式进行评估。 研究结果：针对首次就诊场景，预测疾病进展的最优风险因子组合包括泼尼松用药史、患者的预测百分比DLCO与FVC值、杵状指、世界卫生组织（World Health Organization, WHO）功能分级，以及圣乔治呼吸问卷（St. George's Respiratory Questionnaire, SGRQ）的两个条目：药物不良反应与居家行走时的呼吸困难感。该组合的受试者工作特征曲线下面积（AUC）为0.75（95%置信区间：0.7–0.79），模型的灵敏度、特异度与准确率分别为61%、79%与70%。针对随访就诊场景，预测因子包括原始DLCO与FVC的6个月变化量、预测百分比DLCO、视觉模拟评分法评估的咳嗽严重程度、肺动脉高压，以及SGRQ的两个条目：药物不良反应与公众场合不适感。该模型的预测性能有所提升，AUC达到0.813（95%置信区间：0.78–0.85），灵敏度、特异度与准确率分别为75%、72%与73%。 创新贡献：本研究采用数据驱动的方法筛选疾病进展标志物，确定了一套可被临床从业者常规追踪的特征组合，用于识别患者未来12个月的疾病进展概率，可为临床政策制定、实践应用与相关研究提供创新参考。 研发说明：本应用采用R统计软件编写，并依托R Shiny包完成部署。研究数据来源于澳大利亚IPF注册研究。