Background and Aims

We aimed to develop deep learning models for the prediction of the risk of advanced colorectal neoplasia (ACRN) in asymptomatic adults based on which colorectal cancer screening could be customized.

Methods

We collected data on 26 clinical and laboratory parameters including age, sex, smoking status, body mass index, complete blood count, blood chemistry, and tumor marker from 70,336 first-time screening colonoscopy recipients. For reference, we used a logistic regression (LR) model with nine variables. Two deep neural network (DNN) models were developed: Model 1 used the same nine variables and Model 2 included all 26 variables. The area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were compared for the models in a randomly split test set.

Results

Compared with the LR model (AUC, 0.721; 95% confidence interval [CI], 0.680–0.762), the DNN Model 1 (AUC, 0.817; 95% CI, 0.789–0.847) and Model 2 (AUC, 0.860; 95% CI, 0.837–0.883) showed significantly improved performance with respect to the prediction of ACRN (both P<.001). For a sensitivity of 80%, the specificity of the LR, and DNN Models 1 and 2 were 50.5%, 65.8%, and 78.8%, respectively (both P vs LR model < .001), indicating that the colonoscopy workload required to detect a similar number of ACRNs could be reduced by 58.4% and 73.7% using the DNNs.

Conclusions

The application of DNNs to big clinical data could significantly improve the prediction of ACRNs compared with the LR model, implying the potential for realizing further improvements using large quantities and various types of biomedical information.

中文翻译：

---

应用深度学习利用临床大数据预测晚期肿瘤在无症状成人结直肠癌筛查中的应用

 背景和目标

我们的目标是开发深度学习模型，用于预测无症状成人晚期结直肠肿瘤 (ACRN) 的风险，并在此基础上定制结直肠癌筛查。

 方法

我们从 70,336 名首次结肠镜检查接受者收集了 26 项临床和实验室参数的数据，包括年龄、性别、吸烟状况、体重指数、全血细胞计数、血液化学和肿瘤标志物。作为参考，我们使用了具有九个变量的逻辑回归 (LR) 模型。开发了两个深度神经网络 (DNN) 模型：模型 1 使用相同的 9 个变量，模型 2 包含所有 26 个变量。在随机分割的测试集中比较模型的受试者工作特征曲线下面积 (AUC)、敏感性和特异性。

 结果

与 LR 模型（AUC，0.721；95% 置信区间 [CI]，0.680–0.762）相比，DNN 模型 1（AUC，0.817；95% CI，0.789–0.847）和模型 2（AUC，0.860；95% CI） CI，0.837–0.883）在 ACRN 预测方面表现出显着改善的性能（均P <.001）。对于 80% 的灵敏度，LR 以及 DNN 模型 1 和 2 的特异性分别为 50.5%、65.8% 和 78.8%（ P与 LR 模型均 < .001），表明检测所需的结肠镜检查工作量使用 DNN 可以将类似数量的 ACRN 减少 58.4% 和 73.7%。

 结论

与 LR 模型相比，DNN 在大临床数据中的应用可以显着改善 ACRN 的预测，这意味着利用大量、各种类型的生物医学信息实现进一步改进的潜力。