Division of Health Sciences, Warwick Medical School, University of Warwick, Gibbet Hill Road, Coventry CV4 7AL, UK.
Centre for Cancer Prevention, Wolfson Institute of Preventive Medicine, Queen Mary University of London, London, UK.
Br J Cancer. 2018 Jan;118(2):285-293. doi: 10.1038/bjc.2017.375. Epub 2017 Nov 2.
The faecal immunochemical test (FIT) is replacing the guaiac faecal occult blood test in colorectal cancer screening. Increased uptake and FIT positivity will challenge colonoscopy services. We developed a risk prediction model combining routine screening data with FIT concentration to improve the accuracy of screening referrals.
Multivariate analysis used complete cases of those with a positive FIT (⩾20 μg g) and diagnostic outcome (n=1810; 549 cancers and advanced adenomas). Logistic regression was used to develop a risk prediction model using the FIT result and screening data: age, sex and previous screening history. The model was developed further using a feedforward neural network. Model performance was assessed by discrimination and calibration, and test accuracy was investigated using clinical sensitivity, specificity and receiver operating characteristic curves.
Discrimination improved from 0.628 with just FIT to 0.659 with the risk-adjusted model (P=0.01). Calibration using the Hosmer-Lemeshow test was 0.90 for the risk-adjusted model. The sensitivity improved from 30.78% to 33.15% at similar specificity (FIT threshold of 160 μg g). The neural network further improved model performance and test accuracy.
Combining routinely available risk predictors with the FIT improves the clinical sensitivity of the FIT with an increase in the diagnostic yield of high-risk adenomas.
粪便免疫化学检测（FIT）正在取代粪便隐血试验（FOBT）用于结直肠癌筛查。FIT 检测的广泛应用和阳性率的提高将对结肠镜服务形成挑战。我们开发了一种风险预测模型，将常规筛查数据与 FIT 浓度相结合，以提高筛查推荐的准确性。
多变量分析采用了阳性 FIT（⩾20μg g）和诊断结果的完整病例（n=1810；549 例癌症和高级腺瘤）。使用 FIT 结果和筛查数据（年龄、性别和既往筛查史）进行逻辑回归，建立风险预测模型。进一步使用前馈神经网络对模型进行开发。通过区分度和校准评估模型性能，并通过临床敏感度、特异性和接受者操作特征曲线研究测试准确性。
仅使用 FIT 时，区分度为 0.628，而使用风险调整模型时为 0.659（P=0.01）。Hosmer-Lemeshow 检验校准的风险调整模型为 0.90。在保持相似特异性（FIT 阈值为 160μg g）的情况下，敏感性从 30.78%提高到 33.15%。神经网络进一步提高了模型性能和测试准确性。
将常规可用的风险预测因子与 FIT 相结合，可以提高 FIT 的临床敏感度，并增加高危腺瘤的检出率。
