Machine learning for psychological disorder prediction in Indians during COVID-19 nationwide lockdown

2.1COVID-19

COVID-19, is an acronym for coronavirus disease 2019, where CO stands for corona, VI for virus and D for disease, while 19 is for the year that the outbreak was first identified (31 December 2019) [5]. The International Committee on Taxonomy of Viruses55 refers the virus as “severe acute respiratory syndrome coronavirus 2”, or SARS-CoV-2. The disease was first identified in 2019 in Wuhan, the capital of China’s Hubei province, and on 11 March, 2020, the World Health Organization (WHO) announced that the outbreak should be considered a pandemic as it spread globally across multiple countries. It is a contagious respiratory illness with fever, dry cough and tiredness as commonly reported symptoms. The virus can be asymptomatic or in mild cases people may get just a runny nose or a sore throat. In the most severe cases, people with the virus can develop difficulty breathing, and may ultimately experience a multiple-organ failure causing fatalities. The COVID-19 virus spreads primarily through droplets of saliva or discharge from the nose when an infected person coughs or sneezes [6]. Figure 2 shows the microscopic image of SARS-CoV-2. The spikes on the outer edge of the virus particles resemble a crown, giving the virus its characteristic name.

Figure 2.

Microscopic image of SARS-CoV-2.

There is no available vaccine for COVID-19. Social distancing (stay at home, no handshakes), practicing good hygiene (washing hands) and respiratory etiquette (for example, by coughing or sneezing with a flexed elbow) are some key preventive methodologies adopted to prevent community transmission. The Indian Council of Medical Research (ICMR)66 categories the spread of disease into four stages, namely stage-1 where transmission is limited to cases with travel history to other countries, stage-2 is local transmission from known infected persons, stage-3 is community transmission and stage-4 is when it turns into an epidemic. An epidemic that has spread over several countries or continents, usually affecting a large number of people is referred to as a pandemic. Though, India is at stage-2 now, the government has begun preparing for stage-3 and taken stern precautions to curb social gatherings. A 21-day nationwide lockdown starting 24th March 2020 has been announced for all 1.3 Billion Indians.

2.2COVID-19 and mental health

Though stress and sadness are normal and natural responses to COVID-19, self-isolation, social distancing, quarantine and a national lockdown can all be emotionally challenging. People who are working in different cities and live away from their families, all alone, are now stuck at home for this given time period. Their daily routine of going to office, hanging out with their friends or colleagues and meeting people in public spaces; things that give them joy and keep them entertained has come to a halt. This sudden arrangement has traumatized many people and triggered corona-phobia. The outbreak is continually affecting the global economy and consequently the people are facing job risks or financial insecurity. COVID-19 can add to one’s “typical” levels of stress and worry and can lead to increase in distress, symptoms and relapse into mental health. Further, epidemics lead to stigmatization of affected individuals, authority figures, and health care professionals. Reports regarding threatening by landlords and ostracizing healthcare professionals (doctors, nurses and paramedics) treating COVID-19 patients and aviation professionals (flight crew) have been reported in various news media in India.77 Undoubtedly, this behavior reflects the levels of psychological distress that people are suffering amid the global Coronavirus outbreak. Further, the fear of health contamination may lead someone to become obsessed with hand hygiene and trigger compulsion to wash or clean spiking the obsessive compulsive disorder. The psychological effects can lead to a range of reactions from fear to anxiety, from boredom to depression, from frustration to anger, from stigmatization to repulsion. While it is vital to remain mentally healthy it is equally essential to seek professional support & intervention if symptoms of this Corona-related “Coro-anxiety” impair daily functioning or personal/professional relationships.

2.3Psychological disorders

Feeling of uncertainty and panic are natural when a traumatic or grueling event occurs. But for some people the struggle becomes overwhelming when worries and fears start interfering with relationships and daily life. Formally, mental illnesses are health conditions involving changes in emotion, thinking or behavior (or a combination of these) [7]. The general cognitive function is hindered to an extent that it can trigger inappropriate responses because those responses are based upon inaccurate thoughts. That is, the person finds difficult to stay focused, process information, store it in memory, and accurately respond. Mental illness is conceptualized as a clinically significant behavioral dysfunction or psychological syndrome. There many different categories of mental/psychological disorders defined in the ICD-10, 10th revision of the International Statistical Classification of Diseases and Related Health Problems (ICD), a medical classification list by the World Health Organization (WHO) known as Mental and behavioral disorders, ICD codes F00 to F99. Figure 3 show the broad categories of psychological disorders [8].

Figure 3.

Categories of psychological disorder.

In this work, we considered binary classification into two types of common psychological disorder, namely, anxiety disorder (AD) and mood disorder (MD). The symptoms of AD and MD significantly overlap but mental markers such as fear and worry are commonly associated with AD whereas hopelessness and worthlessness are associated with MD.

2.3.1Anxiety disorder

Anxiety disorder [9] is an umbrella term to pool a collection of psychological disorders characterized by significant feelings of panic and fear (Fig. 4). It is an intrusive apprehension about future and frightfulness to the present situation, making it difficult to concentrate and accomplish daily tasks. It is an excessive, irrational dread of everyday situations which can interfere with daily activities. These feelings may cause physical symptoms, such as headaches, chest pain, palpitations, stomach ache and shakiness.

Figure 4.

Types of anxiety disorders.

Figure 5.

Types of mood disorders.

The following symptoms are some prominent markers of anxiety disorder:

• • Difficulty in concentrating and carrying routine work.

• • Restlessness.

• • Excessive worrying & stress.

• • Irrational phobias (fear triggered by a specific stimulus or situation, like fear of animals, heights, blood, natural events, situations).

• • Significant weight gain/weight loss.

• • Altered sleep pattern (insomnia, hypersomnia, parasomnia-sleep terrors, sleep walking, sleep talking, nightmare).

• • Changed eating habits (anorexia or binge-eating).

• • Separation fear.

• • Adjustment difficulties.

• • Obsessions – unrealistic thoughts or mental impulses.

4.1Participant dataset creation and feature extraction

Generically, there are four different data collection techniques – observation, questionnaire, interview and focus group discussion [39]. In this work, the dataset was prepared by evaluating the responses of an online mental health questionnaire. The response were recorded from 26th March 2020 to 1st April 2020. The questionnaire was prepared with 12 closed-ended questions using Google forms and was circulated online for voluntary participation. The confidentiality of information was assured. The participants were asked to ‘select all that apply’ for each response. A total of 460 participant response sheets were received out of which 65 participants opted for the response “The current Corona crisis does not affect me” and were excluded from the study. Consequently, 395 participant response sheets were considered for to study the impact of COVID-19 lockdown and to effectively detect the type of psychological disorder during this restriction. As the study was focused on Indians, we included participants who are residing in India and Indians who are currently living abroad (Non- Resident Indians, NRI). The intent was to examine the mental-health conditions of both domestic & international Indians during this pandemic. The questionnaire included self-reported demo-graphic data about gender; occupation (government job, private job, self-employed, housewife, student, retired, special/essential service providers-doctors/para-medics/flight crew); age (16–24, 25–60, 61–100) and marital/partnership status (married, unmarried, separated). The number of NRI participants accounted for 18.23% (72 participants) in comparison to 81.77% (323 participants) domestic respondents. The gender, age and occupation distribution collected from the questionnaire is shown Figs 7–9 respectively.

Figure 7.

Gender distribution.

Figure 8.

Age distribution.

Figure 9.

Occupation distribution.

Generally, to measure depression a benchmark Patient Health Questionnaire, PHQ-9 [40] which consists of nine symptomatic questions to detect depression and its severity is used. But as the intent of this work was to identify the onset of specific psychological disorders, namely anxiety and mood disorders during this corona crisis, a COVID-19 Mental Health Questionnaire (MHQ COVID-19) is designed. This web-based questionnaire consisted of questions characterizing 18 prominent symptomatic markers for the respective disorders. As discussed earlier, there is a significant overlap in the symptoms of both the disorders but the cumulative results of symptoms were used to predict the final class of disorder. The following Table 1 depicts the symptomatic queries with their respective descriptions and disorder categories:

Each symptom is used as feature to train the classifier which can predict the disorder type. The feature vector is created using these 18 symptomatic markers where the feature value is set to ‘1’ on the presence of symptom, else it is set to ‘0’.

4.2Binary classification using supervised machine learning

Supervised learning is the learning process when the outcome variable is known, and that this information is explicitly used in training (supervised). Classification in machine learning (ML) is the task of segregating instances of input into various classes based on previous training input [41]. It is a supervised task, wherein a number of labelled training instances are given to the ML classifier as input. The classifier learns from the training sample. Then the performance is tested on a sample of test instances.

In this work, a multinomial Naïve Bayes Classifier [42] is used. Naive Bayes is a family of probabilistic algorithms that take advantage of probability theory and Bayes’ Theorem to predict the category of a sample. It predicts a class by taking strong independence assumptions over features/predictors. Every feature/predictor (x⁢_⁢i) is conditionally independent of its class (c). The probability of a class belonging to features is given by Eq. (1):

where, c is the Class X={x1,x2,x3,…,xn} are the set of features.

A naive Bayes classifier which uses a multinomial distribution for each of the features is referred to as a Multinomial Naive Bayes. NB is an easy to implement, simple, scalable and a fast technique. It is a generative model as compared to logistic regression (LR), decision trees (DT) or support vector machines (SVM) which are discriminative. With the small training data, LR and DT model estimates may over fit the data. An interesting finding is that NB can perform better when there is low amount of training data. Another interesting characteristic of NB is that, when computational time is a critical factor these classifiers may be quickly learned from a dataset by means of the original inductors.

Figure 10.

Distribution of symptomatic markers in MHQ COVID-19.

5.1MHQ COVID-19 outcomes

The result of survey clearly indicate the adverse effects on mental health because of lockdown for infection prevention. Almost 86% of the participants had self-reported symptoms of anxiety and depression. The distribution of symptomatic markers is shown in Fig. 10. The top five in descending order of distribution were altered sleep habits, boredom, altered eating habits, loss of freedom and lack of interest in activity previously enjoyed.

The separation of 395 records into anxiety and mood disorder types was 60.75% for AD and 39.24% for mood. It was also observed that the number of females with MD were more in comparison to males. Figures 11 and 12 show the Male-female case distribution for MD and AD respectively.

Figure 11.

Male-female distribution for MD.

Figure 12.

Male-female distribution for AD.

It was also noted that out of 80.41% male with AD, nearly 50% were in private job and 20% were self-employed where as in case of MD out of 61.94% females around 42% were housewife followed by 26% in private job. The student category had the minimal number of AD cases. Also, out of the 45 female and 27 male NRIs the distribution of MD and AD in females was 53.4% and 46.6% respectively, whereas NRI men had nearly equal distribution of disorders. The results are clearly indicative of mass onset of Coro-anxiety and the need of early intervention to immobilize this big silent killer.

5.2Performance of PDP tool

The proposed PDP tool was evaluated for classification performance using to ROC-AUC curve. The participant dataset of 395 records was labelled for anxiety and mood disorder. It comprised of 240 records labelled as AD and 155 labelled as MD. The classification results were assessed by partitioning the dataset into training and test sets. 10-fold cross-validation was performed to create a validation set and find the best parameters. The training has been performed on 80% of the dataset. The prediction results are shown using the confusion matrix (Fig. 13). It helps to visualize the proportions between the predicted and true label for both the datasets to calculate the performance of the developed model using some evaluation metrics.

To evaluate the performance of the classifier, we calculated precision, recall and accuracy [43], where

Figure 14 depicts the confusion matrix of the PDP classifier. An accuracy of 92.15% was achieved. The recall and precision values were 0.9485 and 0.9208 respectively.

The ROC-AUC curve of the proposed PDP tool is shown in Fig. 15. ROC stands for Receiver Operating Characteristic. ROC is a probability curve which is created by plotting the true positive rate (TPR) against the false positive rate (FPR) at various threshold settings. Area under the ROC curve (i.e. AUC) tells how much model is capable of distinguishing between classes. The model that has higher TRP for each point of FPR will give better performance, that is higher the AUC, better the model is at predicting true class labels with max probability.

Figure 15.

ROC-AUC curve of PDP tool.

A comparative analysis with four other classification algorithms, namely, Decision Tree (DT), Support vector machine (SVM), k-nearest neighbor (K-NN) and Logistic Regression (LR) was also done to compare it with the Naïve Bayes Classifier (NB). It was observed that NB outperformed all the other classifiers in terms of accuracy (Fig. 16).

Figure 16.

Accuracy of classifiers.

The comparative analysis of the classification algorithms is given in Table 2.

The least precision obtained was for the DT classifier and correspondingly a lower accuracy tree. This is because tree splitting is locally greedy and therefore DT is more likely to get stuck in local optima.