Hemoglobin A1C
Last updated
Last updated
Smartphone-based Identification of Critical Levels of Glycated Hemoglobin A1c
Abstract
A current health concern is the constraints of blood glucose monitoring techniques in the face of the ever-expanding predominance of diabetes. Electronic medical devices can potentially overcome these limitations and prevent the development of diabetes-related complications. This study investigated whether advanced machine learning methods, a smartphone-based AI Imaging Technology that assesses health markers, can be a viable solution for diabetes management. To examine the validity and a novel machine algorithm for diabetes prediction, we compared the diabetes classification from AI Imaging obtained glycated haemoglobin A1c (HbA1c) concentrations against data obtained from FDA-approved blood immunoassays. The data set was obtained from 513 participants recruited during their annual physical examination at the Health Management Centre of the Affiliated Hospital of Nanjing University Medical School, China. We used a kitchen sink random forest machine algorithm for diabetes prediction. To validate the model, pristine testing was done on 400 pristine participants pseudo-randomly selected during 20 trials of training and testing. The confusion matrix found BitDoctor AI imaging Technology to have a classification accuracy of 66%, and the Receiver operating characteristic (ROC) curve of the Random Forest (RF) classifier found AI Imaging to have a ROC Area Under the Curve (AUC) of .69. The present study provides evidence for the potential use of the technology for contactless, non-invasive, and inexpensive assessments of diabetes.
The BitDoctor AI imaging Technology system, the smartphone application, was used to collect participants’ facial blood flow information. BitDoctor AI imaging Technology to detect and track an individual’s facial regions of interest (ROIs), which provide the optical properties of the face to obtain blood flow information.
Machine Learning Analysis.
The dataset was divided into two parts: nondiabetes and diabetes. A novel kitchen sink random forest model to utilize the facial blood flow information obtained from AI imaging for diabetes classification. Since HbA1c accounts for 97% of total hemoglobin, a constructed a model to particularly extract HbA1c from AI Imaging obtained hemoglobin concentration (Kahn & Fonseca, 2008). Thus, the video of each participant's face was analyzed for facial blood flow information that reflects cardiovascular activities that correlate with HbA1c. In order to construct the Random Forest (RF) model, MATLAB model to build not balanced decision trees for ensemble diabetes prediction. Random samples of the training dataset were used to build the not-balanced trees. Each decision tree was built to make independent diabetes classification and ‘vote’ for the corresponding class (i.e., non-diabetes or diabetes) based upon HbA1c thresholds. In addition, MATLAB bagging and feature randomness methods to produce uncorrelated decision trees. Also applied the MATLAB feature selection and not-balanced RF classifier training with a threshold for non-diabetes and diabetes equal to clinically approved ones, 5.7% (Sherwani et al., 2016). The results from the laboratory HbA1c test were used as ground truth data in order to evaluate the accuracy of our diabetes classification model.
Validation of the Model. To validate the model, data were derived from pristine testing on 400 pristine participants pseudo-randomly selected during 20 trials of training and testing. To compare the data obtained from AI Imaging against those from the blood samples, the data computed the MATLAB confusion matrix and ROC curve of RF classifier to assess the level of agreement between AI Imaging and the blood samples’ diabetes prediction. The confusion matrix was used to assess the number of false positives, false negatives, true positives, and true negatives of BitDoctor AI Imaging. The ROC curve of RF classifier was used to assess the true positive rate and false positive rate of BitDoctor AI Imaging. The threshold for diabetes was incorporated in the MATLAB feature selection and not-balanced random forest classifier training at 5.7%
The ROC curve of the RF classifier is a graphical illustration of a model’s prediction of binary outcomes. It was used to assess the true positive rate and false positive rate of BitDoctor AI Imaging.
The confusion matrix. There are two axes: (1) the true class (y-axis; determined by blood samples), and (2) the predicted class (x-axis;). Non-diabetes is classified as ‘0’, and diabetes is classified as ‘1’. The equal error rate (EER) threshold values for the model’s false positive and false negative rate is set to {159 100 55 86 } 0.351. BitDoctor Technology has a classification accuracy of 66%.
