The usefulness of machine learning analysis for predicting the presence of depression with the results of the Korea National Health and Nutrition Examination Survey.

BACKGROUND

Depression is a major public health concern, with an estimated 10.8% of adults experiencing depression. Depression can have a significant impact on an individual's quality of life, social function, and productivity. Early diagnosis of depression is important in preventing its progression. Several tools, such as the Patient Health Questionnaire-9 (PHQ-9) and Beck Depression Inventory, are used to screen patients for depression. We investigated the potential of machine learning in predicting the presence of depression using the results of a national survey.

METHODS

We collected the data of 5,420 patients from the 2020 Korea National Health and Nutrition Examination. The presence of depression was defined as ≥5 PHQ-9. We categorized output variables into the presence of depression (PHQ-9, ≥5) and absence of depression (PHQ-9, <5). We used 20 variables related to sociodemographic characteristics, health behavior, and presence of chronic disease for the development of three machine learning algorithms [random forest, logistic regression, and deep neural network (DNN)]. Eighty-seven decision trees were used for the random forest model. Linear regression algorithm shows a linear relationship between various input and output variables. For the DNN model, three layers with 16-32-64 neurons, Adam optimizer, and rectified linear unit (ReLU) activation were used. Of the included samples, 70% and 30% were randomly divided into the training and test sets, respectively.

RESULTS

The area under the curve (AUC) of the test dataset for the random forest model was 0.803 [95% confidence interval (CI), 0.776-0.829], 0.812 (95% CI, 0.787-0.837) for the logistic regression model, and 0.805 (95% CI, 0.780-0.831) for the DNN model.

CONCLUSIONS

Our study demonstrated the potential of machine learning for the development of models for predicting the presence of depression based of various health-related data. Machine learning models can potentially overcome the limitations of traditional diagnostic methods for depression by incorporating a wide range of objective variables to accurately identify patients with depression, thus avoiding the subjectivity and potential diagnostic errors associated with the subjective interpretation of symptoms observed by a clinician. Further efforts to increase the accuracy of machine learning models by utilizing more variables and data needed to detect depression.