Prediction Model of Metabolic Syndrome in Non-obese Body Population
Metabolic syndrome is defined as a cluster of conditions that occur together that increases; the risk of certain diseases. Based on data, a person is diagnosed to be metabolic syndrome positive if s/he has three or more of the conditions below (Fig. 1):
- Abdominal obesity, measured by waist circumference greater than 40 inches for men, or greater than 35 inches for women
- Triglyceride level of 150 milligrams per deciliter of blood (mg/dL) or greater
- HDL(high-density lipoprotein) cholesterol of less than 40 mg/dL for men or less than 50 mg/dL for women
- Systolic blood pressure (top number) of 130 mmHg or greater, or diastolic blood pressure (bottom number) of 85 mm Hg or greater
- Fasting glucose of 100 mg/dL or greater
Metabolic syndrome is a serious health condition that affects about 23 percent of adults. Metabolic syndrome positive persons are at higher risk of cardiovascular disease, diabetes, stroke and other diseases.
Metabolic syndrome positiveness is typically expected to be found for obese people. However, there is another population of people who are not obese but metabolic syndrome positive. Usually, these non-obese but metabolic syndrome positive people are unaware of their condition and may think that they are healthy.
In the project, we wanted to focus on these non-obese but metabolic syndrome positive people and develop a prediction model to predict metabolic syndrome positiveness/negativeness in non-obese body population on the basis of demographic and environmental factors.
Fig. 1. Decision of of metabolic syndrome and its effects on diseases
Our EHR (electronic health records) example is shown in Table 1. With such EHR, we did data processing through machine learning classifications as in Fig. 2.
In Table 1, we can see that there is significant output class imbalance. We solve the output class imbalance problem by applying an oversampling method called SMOTE(synthetic minority oversampling technique) to the majority class records of the training set, followed by applying random downsampling to the minority class records of the training set. However, positive:negative ratio of 1:1 on the oversampled/downsampled training set may cause significant differences between the distribution of the training set and the distribution of the test set, overfitting problems may occur. As a solution, we chose not to aim at positive:negative ratio of 1:1. In other words, after all these oversampling/downsampling, we still have class imbalance problems.
As in Fig. 2, since we are focusing only on non-obese population in this research, we selected only population satisfying BMI(body mass index)< 25km/m2. After handling missing categorical variables, selecting for BMI < 25km/m2, followed by oversampling/downsampling, we finally got 30953 records on the training set, and 17514 records on the test set.
Table 1. An EHR example
Fig. 2. Overall flow of data processing ~ classification
Since we are considering logistic regression as one of our machine learning algorithms in our research, we also checked whether there is a linear relationship between the logit of the outcome and each of the predictor variables, as in Fig. 3.
Fig. 3. Linear relationship checked between the logit of the outcome and each of the predictor variables.
Besides, we also checked whether there is little or no multicollinearity among the predictor variables. We did multicollinearity checks separately for the categorical variables and continuous variables, respectively. We find that in particular, there is significantly high multicollinearity in the continuous variable plot. This can be understood, since BMI is indeed calculated from height and weight. Plus, height and weight are usually in a linear relationship.
Fig. 4. Multicollinearity results
In efforts to decrease the high multicollinearity among the continuous predictor variables, we adopted three different models as in Fig. 2, where model A considers BMI, model B considers both height and weight, and model considers all of BMI, height, and weight. For all of model A ~ C, age, sex, smoking, alcohol, exercise were commonly included.
Our machine learning modelling results can be found in table 2. As it is common to get results with high accuracy but low sensitivity(or high accuracy but low specificity) in typical class-imbalanced problems, our parameter tuning was performed in the domain of F1-score. In addition, we checked performance for accuracy, precision, sensitivity, and F1-score. Our results in table 2 shows that both logistic regression and random forest classification are able to give alerts to metabolic syndrome positive persons.
Table 2. Results from applying machine learning algorithms
Findings and Future Work
In this research, sing basic health checkup results and machine learning, it is possible to to predict metabolic syndromes positiveness with high accuracy/sensitivity/specificity/precision/f1-score.
In the future work, we are planning to write an R Shiny app which tells predictions for metabolic syndrome positiveness/negativeness to anyone interested, based on his/her basic information such as height, weight, etc, input to the R Shiny app.
* Under submission for a journal