With the increasing impact of artificial intelligence in the domain of healthcare, unsupervised learning techniques have been in discussion for the detection of hidden patterns or relationships between diferent data features. The domain of mental health is no exception. This paper presents an unsupervised machine learning approach to analyze the presence of depression and anxiety. It presents unsupervised models to analyze depression based on depression-related data of a student cohort containing demographic and academic data along with depression information collected through the Beck Depression Inventory questionnaire, in addition to scores such as the PHQ (Patient Health Questionnaire) score, GAD (Generalized Anxiety Disorder) score, and Epworth score, which provide insights into the severity and impact of depressive symptoms, anxiety symptoms, and daytime sleepiness, respectively. The methodology involves data collection and preparation, feature selection, modeling, anomaly detection, and analysis of the relationships between diferent features.
Depression, or depressive disorder, is a common mental health disorder that afects around
4% of the world population1. It can afect people of any age and its main symptoms are mood
changes, loss of interest in daily activities, and sleep and appetite alterations among others2.
It was studied that mood disorders, including diferent types of depression, are one of the
most common comorbidities found among people who commit suicide [
Unsupervised learning makes sense when only the input data is known, and no output data
corresponds to a given input. That is, only the structure of the data can be described, to try to find
some pattern that simplifies the analysis. Clustering looks for groupings based on similarities,
but nothing guarantees that these have any meaning. When discussing clustering techniques,
one of the most popular techniques is the K-Means [
In this context, the objective of this paper is to use unsupervised modeling to analyze if the cohorts could be separated into diferent groups with distinguishing risk factors of the depression and anxiety dataset named "Depression and anxiety data" built by Shahzad 5 which focuses on college students. Although a dataset presents labeled data, the study does not use it because it aims to find hidden relationships among their values to gain better insights about depression and/or anxiety diagnoses. Besides this primary goal, a secondary objective is to analyze the relationships between the features and their impact on the development of distinguishing risk profiles in the cohorts.
The final goal of such research is to generate a model that could be included as a part of
a decision support system capable of diagnosing mental health diseases such as depression
or anxiety. A Clinical Decision Support System (CDSS) is a software tool that is designed to
help healthcare professionals make decisions by giving personalized information about their
patients[
5Depression and anxiety data. https://www.kaggle.com/datasets/shahzadahmad0402/ depression-and-anxiety-data lines of work.
To achieve the proposed objective the programming language Python and its set of machine learning libraries are used, applying unsupervised learning modules of clustering and anomaly detection over the mentioned dataset. The mentioned dataset contains 783 records. Their characteristics are described below in Table 1.
To eliminate bias, and find the relationship between the diferent characteristics, the columns
"Depression diagnosis" and "Anxiety Diagnosis" were eliminated from the dataset for both
kinds of analyses. The data gathered here include age and body mass index information along
with results from several psychological questionnaires: the Beck Depression Inventory6 (a
standardized method used to determine depression levels), the Beck Anxiety Inventory7 (an
analogous technique used to detect anxiety), the Patient Health Questionnaire (PHQ) [
The process continues with the data preprocessing, which includes the elimination of null and duplicate values, and finally, with the data normalization using the Standard Scaling technique. Then it continues with the clustering process described in the following section, and, finally, the introduction of anomaly detection methods to see if it improves clustering results.
Clustering is one of the most important areas within the unsupervised machine learning domain.
The purpose of this process is to achieve the grouping of unlabeled data sets and to build subsets
of data known as clusters. This way, each cluster within a network is formed by a collection
of data similar to each other but having diferential elements concerning others. This type of
process is applied to unsupervised models, allowing data to be segmented into groups of similar
dimensions. After obtaining the pre-processed dataset, a technique is used to determine the
optimal number of clusters that will eventually be assigned using the K-Means algorithm. The
method used to determine the optimal amount of clusters is the Silhouette, which measures the
quality of clustering, using the separation distance between them. It tells us how close each
point in a cluster is to points in neighboring clusters. This distance measure is in the range
[
6Beck Depression Inventory (BDI). https://www.apa.org/pi/about/publications/caregivers/practice-settings/ assessment/tools/beck-depression 7Beck Anxiety Inventory. https://www.zotero.org/google-docs/?yIRpC0
Characteristics
After getting the optimal number of clustering and dividing the data into diferent clusters.
As mentioned above, the K-Means algorithm is used to achieve such a goal. Being one of the
most powerful and popular methods for applying and understanding the grouping process [
After the data is grouped in diferent clusters, the target variables are introduced again. For each one of the clusters generated from both methods, the T-Test8 is performed to determine if the data present significant diferences among the diferent groups.
The clustering process is performed again, iteratively. The goal is to get the majority of depression and anxiety-diagnosed patients inside a single cluster. However, the process can ifnish sooner if some other conditions are met: if the unhealthy patients start to be separated among diferent clusters, or if the metrics for the new split do not show an improvement. This process, in case it is needed, is referred to as sub-clustering. Figure 1 summarizes the Methodology pipeline. After one cluster satisfies the conditions mentioned above, the last step is to analyze the results obtained by these processes.
As explained above, Anomaly Detection is a technique for identifying events that are statistically diferent from the rest. Such anomalous behavior is usually translated as a problem. Anomaly Detection has a diferent standpoint, based on models and data predictions (see Figure 2).
Anomaly Detection is used for the detection of patients whose features do not match with the majority of the population, allowing analyzing them diferently. This analysis can lead to seeking better metrics for the predictor models. Two groups are considered: the normal data (those values that are in match with the majority) and outliers (the anomalous values) and they are determined for each feature from the dataset, this way, it can be obtained the conclusion on the behavior of each feature. For each numeric or categorical feature, it is determined the range of normal values and its average. Also, there are identified the upper and bottom anomalous values closest to the normal ones. Besides, to improve the analysis, the average values for superior and inferior anomalous values are calculated. Taking into account the concept of 8KoreaMed Synapse. https://synapse.koreamed.org/articles/1156170 clustering, it can be represented with the following graph the situation in which certain data are grouped in several clusters, leaving out the so-called outliers.
Besides the detection of anomalous values, the process also includes analyzing the relationship between anomalous values from diferent features. Seeing whether anomalous values for a patient in a specific feature are also reflected in another diferent feature may help determine how those diferent features behave and find relationships between them. Discovering these relationships makes it possible to further justify decision-making on how to act with that patient. These are decisions that will later afect the tuning of the predictive model.
Following a similar presentation to the previous section, the results for both approaches, Clustering and Anomaly Detection are presented.
Before determining the optimal number of clusters, the dataset is normalized using the Standard Scaling method. Then, as mentioned in the Methodology section, the process continues by trying to obtain the optimal number of clusters. For each one of the cluster combinations, it is observed the diferences between the cluster values and the Silhouette average value for that amount of clusters. For all the combinations, except the 9-cluster one, all the values are above the Silhouette average value, meaning that they are good candidates for the division.
Taking into account the Average Silhouette Width (ASW), which indicates the degree of similarity of an item in a cluster compared to others, the value ranges from -1 (the sample is not very similar to its neighbors and little diferent from those of other clusters) to 1 (the sample is very similar to its neighbors and very little similar to those of other clusters). According to the Silhouette analysis, the convenient number of clusters is selected in 2 to group the data, this being the one with the highest value, considering a good cluster configuration. However, the 2-clusters division is decided because its average value is bigger.
Then, the T-Test method is used to check whether the diference between the clusters is significant enough. For the 2-cluster split, 78.95% of the P-Values for all the features are less than the selected threshold value (0.05), including 100% of the target variables, showing an important diference among the clusters, especially in the target variables which are the ones this procedure are trying to separate. Using that information, the 2-cluster division is concluded to separate the data efectively.
The data obtained for the diferent clusters are described following in Table 2. Table 3 shows how the target variables, once introduced again into each cluster, behave diferently. Both features, in this case, fulfill the conditions analyzed by the T-Test.
It is observed that the first cluster (Cluster 0) groups all the patients with depression and anxiety levels above normal, while the second cluster (Cluster 1) tends to contain healthy patients. However, the first cluster also includes many healthy patients: it doubles the size of the second cluster (531 vs 242) while the original analysis visible in Table 1 presents only 10% of the cohort as diagnosed with one of the diseases. So, to diferentiate unhealthy patients the same methodology (scaling, number of cluster determination, T-Test, clustering, and analysis) is applied again but only over this first cluster.
The Silhouette analysis indicates a value of 4 according to the obtained values for each possible number of clusters in conjunction with the chart analysis (value of 0.184).
The T-Test analysis is performed later which gives better significant diferences between the features of the diferent sub-clusters in the 4-cluster combination (63.16%), but on the target variables, the test is passed in only 16.67% of the clusters. However, Table 4 shows how the sub-clusters are left having in Sub-cluster 1 a slightly larger proportion of unhealthy patients compared with the other groups. The T-Test analysis shows that Sub-cluster 1 has a 33.33% significant diference among the target variables with the other sub-clusters, rising to 66.67% in the case of depression diagnosis.
Table 5 shows how the target variables behave for each one of the sub-clusters once reintroduced. Again, Sub-cluster 1, groups the majority of unhealthy patients, being more significant in the depression diagnosis.
According to this grouping, and taking rows assigned to the Sub-cluster 0 as those identified as positive, a confusion matrix can be defined and, from there, determine the most popular metrics as it was a classification problem. Table 6 shows this.
Although the first attempt to group properly unhealthy patients, it creates a big set of healthy and unhealthy patients, leaving a small group of healthy patients in the smaller set, while the proportion in the dataset is exactly the opposite. When attempting to divide that second and larger cluster to identify better unhealthy patients, a minor sub-cluster is obtained which tends to group them, but the diference with the other sub-clusters is not so significant. The iterative process is stopped at this moment because the next iteration started to show that the unhealthy patients were not grouped into a single cluster anymore. *Features considered as significant according to the T-Test analysis.
The Anomaly Detection analysis gives the results presented on the charts enumerated in Figure 3. School year, given it is a very balanced feature, does not show anomalous values, while the rest of the features do. Figures 3(a) through 3(d) represent features related to the patient body or
Row
Cluster 0 (n=242)
PHQ score Depression Severity
First year: 56% Second year: 43% Third year: 1% 19.19 ± 1.08 Female: 50% Male: 50%
Approach
Accuracy
Recall
Precision
Specificity epression diagnosis Anxiety diagnosis age, while Figures 3(e) through 3(i), which are shown following, indicate mental health-related features.
Most of the features present anomalous values above and below the set of normal values, (a) School Years feature distribution (b) Age feature distribution (c) BMI feature distribution (d) WHO BMI feature distribution (e) PHQ Score feature distribution (f) Depression Severity feature distribution (g) GAD Score feature distribution (h) PHQ Score feature distribution (i) Epworth Score feature distribution being the exceptions the features Age and Anxiety Severity, whose anomalous values are only above the normal data. Table 7 summarizes the number of normal data (normal values) and outliers (anomalous values) for all these features.
Also, binary columns with imbalanced data, such as Suicidal, Depression Treatment, and Anxiety Treatment are considered anomalous when the value is positive.
On the other hand, it was analyzed how diferent anomalous values relate among the diferent features. Table 8 shows these behaviors.
It can be observed that in each case that the patient is detected as anomalous for the PHQ Score feature, it is also detected as anomalous for the feature Depression Severity and vice-versa. Something similar happens between GAD Score and Anxiety.
The clustering analysis returns good accuracy and specificity values in contrast with poor recall and precision values. This suggests that this kind of classification would be very helpful excluding healthy patients but still won’t be able to precisely detect unhealthy patients. It is also observed many similarities between the depressive and anxious patient cohorts, giving the hint that both disorders share similar aspects. Anomaly Detection analysis informs which patients may be altering the dataset so it can be part of future works on the usage of this kind of algorithm to decide what can be done with such patients for the classifier model to improve its metrics: they can be removed or analyzed separately. However, that analysis must be done on a feature level because some anomalous values could be very important to be analyzed as part of the dataset because they may indicate the presence of the disease, especially in such imbalanced data.
Another benefit that can be derived from the anomaly detection process is to find which patterns the anomalous records present to use them to classify the dataset. In this article, Anomaly Detection is used to analyze categorical and numerical features and their relationships, but in the future lines, the same analysis can be performed at a record level.
In conclusion, future open lines of work are mentioned, which include opening the same analysis to other datasets, and, after enough validation, the inclusion of these techniques in decision support systems that help mental health professionals to diagnose better and faster this kind of diseases. Also, the iterating process mentioned in the clustering process can be extended by making more indulgent the stop criteria and checking if that helps the detection of unhealthy patients.
As mentioned above, all this analysis, both the clustering and the anomaly Detection, are meant to be part of a Clinical Decision Support System pipeline, where both parts can help identify patients with diseases or at least anomalous, that may require a deeper inspection.
This work was supported and financed by the Cloudgenia group through its technical and operational initiatives.