Hypercholesterolemia

Description

Hypercholesterolemia risk is the likelihood that the user has abnormally high cholesterol levels (defined as a total cholesterol (TC)-to-high density lipoprotein (HDL) cholesterol (“good cholesterol”) ratio of 4.1 or higher) and corresponds to the percentage of people with the user’s risk profile that have an abnormally high TC/HDL ratio. BitDoctor.ai determines the user’s risk of hypercholesterolemia based on blood flow patterns and demographic information.

Participants

7,000 adults (18+ years of age) recruited at a hospital health clinic in China.

Additional Data Collection

Subjects also received a blood test at the same clinic visit where their cholesterol levels were measured.

Modeling Approach

Blood flow signal was extracted and processed from facial video, and then blood flow features were extracted from blood flow signal.

Feature selection was carried out on blood flow and demographic features to identify features predictive of hypercholesterolemia. A machine-learning based classifier was then created to predict whether an individual has hypercholesterolemia based on these features.

The model was created/trained using 5,600 subjects. The distribution of sex and hypercholesterolemia status in this group is summarized in Table 17.

Subjects in videos	Composition
Sex distribution	46.5 % Male; 53.5% Female
Hypercholesterolemia distribution	30%

Table 17: Hypercholesterolemia distribution of training set

Model Performance

The Hypercholesterolemia Risk model was then validated for accuracy on an independent portion of the dataset (n=700) that was not used in training - validation set. Then the final accuracy was obtained on an independent portion of the dataset (n=700) that was not used in training - test set. The distribution of sex and hypercholesterolemia status in these groups is summarized in Table 18.

Subjects in videos	Composition in Validation	Composition in Test
Sex distribution	47.8% Male; 52.2% Female	46.6% Male; 53.4% Female
Hypercholesterolemia distribution	30%	30%

Table 18: Hypercholesterolemia distribution of validation and test sets

The accuracy on the test set calculated as area under the curve was 80.3% as shown in Figure 22.

Figure 22: AUC of Hypercholesterolemia risk prediction

Predictions are displayed to the user as a percentage likelihood of having hypercholesterolemia. A percentage of 45% or more suggests a risk of diabetes, while 55% or more suggests high risk of diabetes.