Background

Diabetes is a chronic noncommunicable disease with high incidence rate. Diabetics without early diagnosis or standard treatment may contribute to serious multisystem complications, which can be life threatening. Timely detection and intervention of prediabetes is very important to prevent diabetes, because it is inevitable in the development and progress of the disease.

Objective

Our objective was to establish the predictive model that can be applied to evaluate people with blood glucose in high and critical state.

Methods

We established the diabetes risk prediction model formed by a combined TCM tongue diagnosis with machine learning techniques. 1512 subjects were recruited from the hospital. After data preprocessing, we got the dataset 1 and dataset 2. Dataset 1 was used to train classical machine learning model, while dataset 2 was used to train deep learning model. To evaluate the performance of the prediction model, we used Classification Accuracy(CA), Precision, Recall, F1-score, Precision-Recall curve(P-R curve), Area Under the Precision-Recall curve(AUPRC), Receiver Operating Characteristic curve(ROC curve), Area Under the Receiver Operating Characteristic curve(AUROC), then selected the best diabetes risk prediction model.

Results

On the test set of dataset 1, the CA of non-invasive Stacking model was 71 %, micro average AUROC was 0.87, macro average AUROC was 0.84, and micro average AUPRC was 0.77. In the critical blood glucose group, the AUROC was 0.84, AUPRC was 0.67. In the high blood glucose group, AUROC was 0.87, AUPRC was 0.83. On the validation set of dataset 2, the CA of ResNet50 model was 69 %, micro average AUROC was 0.84, macro average AUROC was 0.83, and micro average AUPRC was 0.73. In the critical blood glucose group, AUROC was 0.88, AUPRC was 0.71. In the high blood glucose group, AUROC was 0.80, AUPRC was 0.76. On the test set of dataset 2, the CA of ResNet50 model was 65 %, micro average AUROC was 0.83, macro average AUROC was 0.82, and micro average AUPRC was 0.71. In the critical blood glucose group, the prediction of AUROC was 0.84, AUPRC was 0.60. In the high blood glucose group, AUROC was 0.87, AUPRC was 0.71.

Conclusions

Tongue features can improve the prediction accuracy of the diabetes risk prediction model formed by classical machine learning model significantly. In addition to the excellent performance, Stacking model and ResNet50 model which were recommended had non-invasive operation and were easy to use. Stacking model and ResNet50 model had high precision, low false positive rate and low misdiagnosis rate on detecting hyperglycemia. While on detecting blood glucose value in critical state, Stacking model and ResNet50 model had a high sensitivity, a low false negative rate and a low missed diagnosis rate. The study had proved that the differential changes of tongue features reflected the abnormal glucose metabolism, thus the diabetes risk prediction model formed by a combined TCM tongue diagnosis and machine learning technique was feasible.

中文翻译：

---

基于舌头特征和机器学习技术的非侵入性糖尿病风险预测模型的建立

背景

糖尿病是一种慢性非传染性疾病，发病率很高。没有早期诊断或标准治疗的糖尿病患者可能导致严重的多系统并发症，这可能危及生命。及时检测和干预前驱糖尿病对于预防糖尿病非常重要，因为这在疾病的发展和进程中是不可避免的。

客观的

我们的目标是建立可用于评估处于高危状态和高危状态的人的预测模型。

方法

我们建立了由中医舌诊与机器学习技术相结合而形成的糖尿病风险预测模型。从医院招募了1512名受试者。经过数据预处理后，我们得到了数据集1和数据集2。数据集1用于训练经典机器学习模型，而数据集2用于训练深度学习模型。为了评估预测模型的性能，我们使用了分类精度（CA），精度，召回率，F1分数，精度-召回曲线（PR曲线），精度-召回曲线下的面积（AUPRC），接收器工作特性曲线（ ROC曲线），接收者操作特征曲线下的面积（AUROC），然后选择最佳的糖尿病风险预测模型。

结果

在数据集1的测试集上，无创堆叠模型的CA为71％，微观平均AUROC为0.87，宏观平均AUROC为0.84，微观平均AUPRC为0.77。在临界血糖组中，AUROC为0.84，AUPRC为0.67。在高血糖组中，AUROC为0.87，AUPRC为0.83。在数据集2的验证集上，ResNet50模型的CA为69％，微观平均AUROC为0.84，宏观平均AUROC为0.83，微观平均AUPRC为0.73。在临界血糖组中，AUROC为0.88，AUPRC为0.71。在高血糖组中，AUROC为0.80，AUPRC为0.76。在数据集2的测试集上，ResNet50模型的CA为65％，微观平均AUROC为0.83，宏观平均AUROC为0.82，微观平均AUPRC为0.71。在临界血糖组中，AUROC的预测值为0.84，AUPRC的预测为0.60。

结论

舌头特征可以显着提高经典机器学习模型形成的糖尿病风险预测模型的预测准确性。除了出色的性能外，推荐的Stacking模型和ResNet50模型还具有非侵入性操作且易于使用。Stacking模型和ResNet50模型检测高血糖症的准确率高，假阳性率低，误诊率低。在检测临界状态下的血糖值时，Stacking模型和ResNet50模型具有较高的灵敏度，较低的假阴性率和较低的漏诊率。研究证明，舌头特征的差异性变化反映了糖代谢异常，因此将中医舌头诊断与机器学习技术相结合所形成的糖尿病风险预测模型是可行的。