Elwenspoek Martha M C, O'Donnell Rachel, Jackson Joni, Everitt Hazel, Gillett Peter, Hay Alastair D, Jones Hayley E, Robins Gerry, Watson Jessica C, Mallett Sue, Whiting Penny
The National Institute for Health Research Applied Research Collaboration West (NIHR ARC West), University Hospitals Bristol NHS Foundation Trust, Bristol, BS1 2NT, UK.
Population Health Sciences, Bristol Medical School, University of Bristol, Bristol, BS8 2PS, UK.
EClinicalMedicine. 2022 Apr 7;46:101376. doi: 10.1016/j.eclinm.2022.101376. eCollection 2022 Apr.
Coeliac disease (CD) affects approximately 1% of the population, although only a fraction of patients are diagnosed. Our objective was to develop diagnostic prediction models to help decide who should be offered testing for CD in primary care.
Logistic regression models were developed in Clinical Practice Research Datalink (CPRD) GOLD (between Sep 9, 1987 and Apr 4, 2021, n=107,075) and externally validated in CPRD Aurum (between Jan 1, 1995 and Jan 15, 2021, n=227,915), two UK primary care databases, using (and controlling for) 1:4 nested case-control designs. Candidate predictors included symptoms and chronic conditions identified in current guidelines and using a systematic review of the literature. We used elastic-net regression to further refine the models.
The prediction model included 24, 24, and 21 predictors for children, women, and men, respectively. For children, the strongest predictors were type 1 diabetes, Turner syndrome, IgA deficiency, or first-degree relatives with CD. For women and men, these were anaemia and first-degree relatives. In the development dataset, the models showed good discrimination with a -statistic of 0·84 (95% CI 0·83-0·84) in children, 0·77 (0·77-0·78) in women, and 0·81 (0·81-0·82) in men. External validation discrimination was lower, potentially because 'first-degree relative' was not recorded in the dataset used for validation. Model calibration was poor, tending to overestimate CD risk in all three groups in both datasets.
These prediction models could help identify individuals with an increased risk of CD in relatively low prevalence populations such as primary care. Offering a serological test to these patients could increase case finding for CD. However, this involves offering tests to more people than is currently done. Further work is needed in prospective cohorts to refine and confirm the models and assess clinical and cost effectiveness.
National Institute for Health Research Health Technology Assessment Programme (grant number NIHR129020).
乳糜泻（CD）影响约1%的人口，尽管只有一小部分患者得到诊断。我们的目标是开发诊断预测模型，以帮助确定在初级保健中谁应该接受CD检测。
在英国两个初级保健数据库临床实践研究数据链（CPRD）GOLD（1987年9月9日至2021年4月4日，n = 107,075）中建立逻辑回归模型，并在CPRD Aurum（1995年1月1日至2021年1月15日，n = 227,915）中进行外部验证，采用（并控制）1:4巢式病例对照设计。候选预测因素包括当前指南中确定的症状和慢性病况，并通过对文献的系统综述确定。我们使用弹性网回归进一步优化模型。
预测模型分别包括24个、24个和21个针对儿童、女性和男性的预测因素。对于儿童，最强的预测因素是1型糖尿病、特纳综合征、免疫球蛋白A缺乏症或患有CD的一级亲属。对于女性和男性，这些因素是贫血和一级亲属。在开发数据集中，模型显示出良好的区分度，儿童的C统计量为0.84（95%CI 0.83 - 0.84），女性为0.77（0.77 - 0.78），男性为0.81（0.81 - 0.82）。外部验证的区分度较低，可能是因为用于验证的数据集中未记录“一级亲属关系”。模型校准不佳，在两个数据集中的所有三组中都倾向于高估CD风险。
这些预测模型可以帮助在初级保健等患病率相对较低的人群中识别出CD风险增加的个体。为这些患者提供血清学检测可以增加CD的病例发现。然而，这涉及到比目前更多的人进行检测。需要在前瞻性队列中开展进一步工作，以优化和确认模型，并评估临床和成本效益。
国家卫生研究院卫生技术评估项目（资助编号NIHR129
