Clinicians need to be able to predict which patients will develop specific diseases or complications, and what outcomes are likely in the short or long term. The Framingham and other cohort prediction equations are widely used to predict the risk of coronary heart disease in the next 10 years. However, respiratory medicine does not yet have such widely accepted predictive models. For example, Pinart et al. conducted a systematic review of childhood asthma prediction models. The authors searched for studies of early respiratory symptoms in children of 4 years old or younger as a predictor of asthma between the ages of 6 and 12 years. They identified 14 predictive models mostly using family history, comorbidities or precursors of asthma and severity of early symptoms. Nine models also included clinical tests such as specific immunoglobulin E (IgE) sensitization. However, the predictive ability of all these models was quite poor and there was no single best performing model. Occupational asthma is one of the few preventable phenotypes of asthma in adults. High-risk industries routinely conduct medical examinations of prospective employees, typically including spirometry. Some authorities have advocated the measurement of nonspecific bronchial hyper-responsiveness (BHR) as a part of the pre-employment or pre-placement medical examination. We conducted an inception cohort study in the Australian aluminium industry including a modified Medical Research Council (MRC) questionnaire, spirometry and methacholine challenge test. Baseline BHR was weakly associated with incident chest tightness, but not wheeze or cough over a median of 4 years follow-up. Although the specificity of BHR was 70–77%, the sensitivity was only 24–39% for these asthma symptoms. We concluded that by itself, methacholine challenge was unlikely to be a useful pre-employment or pre-placement screening test in this industry. Another systematic review examined prediction models for the development of chronic obstructive pulmonary disease (COPD). Matheson et al. searched the literature to 2016, selected 30 articles for full-text review, but eventually incorporated only four studies. All of these models included smoking, the strongest known risk factor for COPD, and sex. Some of the models also included age, ethnicity, socio-economic status, height, weight, personal or family history of other lung diseases such as asthma or chronic bronchitis, early life factors, environmental pollution and biomarkers. Perhaps surprisingly, only one model included a measurement of lung function. Some of the models were able to discriminate between those correctly or incorrectly classified as COPD; however, none were particularly accurate at predicting future risk of COPD. Thus, clearly better predictive models are required to be clinically useful for common respiratory diseases such as asthma and COPD. How should the readers or authors of Respirology evaluate or write future publications on predictive models? Fortunately, the editors of 31 journals in respiratory, sleep and critical care medicine have recently published some helpful guidelines. Careful consideration needs to be given to selecting predictor variables, operationalizing these variables, specifying the outcome, dealing with missing data and validating the model, preferably in an external data set. The guidelines describe the metrics available to evaluate the performance of a model. Checklists are also available to ensure that all the key features of a prediction model are accurately reported (e.g. https://www.tripod-statement.org/). We have introduced the statistical considerations for prediction modelling previously in this journal. The new guidelines provide another lens with useful references, and a recent textbook edited by Riley et al. adds substantially to the relevant literature. A key point, highlighted in Figure 1, is that the overall model performance has two distinct aspects: calibration and discrimination. These feed into model utility along with considerations such as disease prevalence or incidence rates, cost-effectiveness of preventive or treatment regimens and so on. We take this opportunity to expand on a point raised by the guidelines, the issues of optimism-corrected internal validation and external validation. Internal validation of a new model is an essential step, confirming its predictive potential with the data source used to develop it. If data are precious, that is, the data source has a limited sample size, then cross-classification is an ideal approach to take for internal validation. The split-sample approach is also valid (if the split is random), but because of its requirement to quarantine a large proportion of the available data, it should only be

中文翻译：

---

呼吸内科预测模型

临床医生需要能够预测哪些患者会出现特定的疾病或并发症，以及短期或长期可能出现的结果。Framingham 和其他队列预测方程被广泛用于预测未来 10 年的冠心病风险。然而，呼吸医学还没有如此广泛接受的预测模型。例如，Pinart 等人。对儿童哮喘预测模型进行了系统评价。作者搜索了有关 4 岁或以下儿童的早期呼吸道症状作为 6 至 12 岁之间哮喘预测因素的研究。他们确定了 14 个预测模型，主要使用家族史、合并症或哮喘的前兆以及早期症状的严重程度。九个模型还包括临床测试，例如特异性免疫球蛋白 E (IgE) 致敏。然而，所有这些模型的预测能力都很差，没有一个表现最好的模型。职业性哮喘是成人少数可预防的哮喘表型之一。高风险行业会定期对潜在员工进行体检，通常包括肺活量测定。一些权威机构提倡将非特异性支气管高反应性 (BHR) 的测量作为就业前或就业前体检的一部分。我们在澳大利亚铝工业中进行了一项初始队列研究，包括修改后的医学研究委员会 (MRC) 问卷、肺活量测定法和乙酰甲胆碱激发试验。基线 BHR 与胸闷事件的相关性较弱，但在平均 4 年的随访中没有喘息或咳嗽。尽管 BHR 的特异性为 70-77%，但对这些哮喘症状的敏感性仅为 24-39%。我们得出的结论是，乙酰甲胆碱挑战本身不太可能成为该行业有用的就业前或入职前筛查测试。另一项系统评价检查了慢性阻塞性肺疾病 (COPD) 发展的预测模型。马西森等人。检索文献至2016年，选取了30篇全文综述，但最终只纳入了4项研究。所有这些模型都包括吸烟、已知的 COPD 最强风险因素和性别。一些模型还包括年龄、种族、社会经济地位、身高、体重、其他肺部疾病（如哮喘或慢性支气管炎）的个人或家族史、早期生活因素、环境污染和生物标志物。也许令人惊讶的是，只有一个模型包括肺功能的测量。一些模型能够区分那些正确或错误地归类为 COPD 的模型；然而，在预测 COPD 的未来风险方面，没有一个特别准确。因此，显然需要更好的预测模型才能在临床上用于常见的呼吸系统疾病，例如哮喘和 COPD。Respirology 的读者或作者应该如何评估或撰写有关预测模型的未来出版物？幸运的是，呼吸、睡眠和重症监护医学领域 31 种期刊的编辑最近发布了一些有用的指南。需要仔细考虑选择预测变量，操作这些变量，指定结果，处理缺失数据并验证模型，最好在外部数据集中。这些指南描述了可用于评估模型性能的指标。还可以使用检查表来确保准确报告预测模型的所有关键特征（例如 https://www.tripod-statement.org/）。我们之前在本期刊中介绍了预测建模的统计注意事项。新指南提供了另一个有用的参考资料，以及最近由 Riley 等人编辑的教科书。大大增加了相关文献。图 1 中突出显示的一个关键点是，整体模型性能有两个不同的方面：校准和判别。这些与疾病流行率或发病率等考虑因素一起进入模型效用，预防或治疗方案的成本效益等。我们借此机会扩展指南提出的观点，即乐观修正的内部验证和外部验证问题。新模型的内部验证是必不可少的步骤，通过用于开发它的数据源确认其预测潜力。如果数据很宝贵，即数据源的样本量有限，那么交叉分类是进行内部验证的理想方法。拆分样本方法也是有效的（如果拆分是随机的），但是由于它需要隔离大部分可用数据，因此它应该只 新模型的内部验证是必不可少的步骤，通过用于开发它的数据源确认其预测潜力。如果数据很宝贵，即数据源的样本量有限，那么交叉分类是进行内部验证的理想方法。拆分样本方法也是有效的（如果拆分是随机的），但是由于它需要隔离大部分可用数据，因此它应该只 新模型的内部验证是必不可少的步骤，通过用于开发它的数据源确认其预测潜力。如果数据很宝贵，即数据源的样本量有限，那么交叉分类是进行内部验证的理想方法。拆分样本方法也是有效的（如果拆分是随机的），但是由于它需要隔离大部分可用数据，因此它应该只