After 7 months of initial reporting, the COVID-19 pandemic continues worldwide. The mental health consequences of the COVID-19 pandemic have been substantial. More than half a million lives and more than 400 million jobs have been lost [1], causing a considerable degree of fear, worry, and concern. These effects are seen in the population at large and may be more pronounced among certain groups such as youth, frontline workers [2], caregivers, and people with chronic medical conditions. The new normal has introduced unprecedented interventions of countrywide lockdowns that are necessary to control the spread but have led to increased social isolation. Loneliness, depression, harmful alcohol and drug use, and self-harm or suicidal behavior are also expected to rise.

The Lancet Psychiatry [3] recently highlighted the needs of vulnerable groups during this time, including those with severe mental illness, learning difficulties, and neurodevelopmental disorders, as well as socially excluded groups such as prisoners, the homeless, and refugees. Calls to action for engaging more early-career psychiatrists [4,5], using technology such as telepsychiatry, and stressing the high susceptibility of frontline medical workers themselves [6] have highlighted the magnitude of the problem. Further, interventions are expected to have a gender-specific impact, with women more likely to be exposed to additional stressors related to informal care, already existing economic disparity, and school closures. Similarly, age and comorbidity status may have a direct impact on susceptibility to mental health challenges due to their relationship with COVID-19 morbidity and mortality. Indeed, it has been established that emotional distress is ubiquitous in affected populations—a finding certain to be echoed in populations affected by the COVID-19 pandemic [7]. Finally, the role of social media [8,9] is complex, with some research indicating an association between social media exposure and a higher prevalence of mental health problems [10].

However, most of these effects have been studied in isolation with a lack of modeling the collective impact of such factors. This study addresses this gap through the use of Bayesian networks (BNs), an explainable artificial intelligence approach that captures the joint multivariate distribution underlying large survey data collected across the United States. We also address the gap of vulnerability prediction for mental health events such as anxiety attacks using supervised machine learning models.

Mental health is a public health concern. Mood disorders and suicide-related outcomes have increased substantially over the last decade among all age groups and genders [21,22]. The rapid spread of COVID-19 forced governments worldwide to close public gathering places, schools, colleges, restaurants, and industries. Social isolation, digital communication, and working and schooling from home have become the new normal, and many jobs have been lost. Collectively, this has triggered a high level of anxiety, stress, and depression globally. We did not find studies that have used models to not only predict but also explain the subtle effects of life situations on mental health. An explainable probabilistic graphical modeling approach with bootstraps and exact inference allowed us to capture many of these effects in a robust manner. Our study revealed that individuals with a prior diagnosis of any mental illness are the most vulnerable for mental illness during the COVID-19 pandemic, which recommends building national-level policies to regularly track their mental status and treat them accordingly. Most importantly, our results reiterate the economic underpinnings of a collective mental health response. Income assistance via Social Security or Supplemental Social Security had a demonstrable effect on the alleviation of anxiety as inferred from our model, which provides the first scientific evidence, to the best of our knowledge, proving the utility of such efforts. The extent of such measures’ effect may be captured in such modeling studies conducted in various parts of the world, with widely varying assistance structures during this time.

Our findings from the United States can also stimulate further cultural and social research in other geographies with similar or different social structures. For example, the effects of in-person communication, as opposed to digital connectedness, may be different in countries where community living and joint families are still commonplace, such as India. Digital connectedness was not as effective as talking to a neighbor, at least in the United States, highlighting that these have fundamentally different influences on mental health and need to be further explored in systematic studies. We conjecture that such differences may arise from the evolutionary mechanisms that have shaped human societies to live and share in close physical connectedness. Such an effect has been previously shown in primates kept in isolation who display depressive symptoms [23,24]. Similarly, parenting and its association with neuropeptide hormones may partially explain [25] our results that the presence of kids reduces anxiety levels. Interestingly, the COVID-19 pandemic has created a unique natural experiment on the collective mental health response of individuals to a health emergency.

The life cycle of such a response may need to be further studied as the world goes through various phases of the pandemic until its resolution. However, our study indicates that the mental health impact is observable within a span of a few months, especially on young individuals. Further research will be needed, ideally in a longitudinal setting, where the same individuals can be surveyed again to understand the dynamics of the collective mental health response.

Our results also highlight that modern technological development in virtual communication is not able to replace natural socializing. Hence, it becomes imperative to design better and more empathetic technological tools that may shape a society and prevent isolation and alienation even while maintaining physical distancing and preventive measures for limiting spread. Personalization and contextualization of such measures will also be important, as our results indicate that persons with previous mental health conditions may be disproportionately affected.

Finally, our results indicate that it may be possible to identify people at the highest risk of developing mental health disturbances. Our model achieved its best performance for those who were most vulnerable (having mental stress more than 5 days in a week) versus least vulnerable (having no stress or less than 1 day of stress in a week). This can help in the segmentation of vulnerable populations such as frontline health care workers and those who are facing disproportionately higher levels of stress during this time.

A key factor in clinical and public health models is transparency and explainability in the face of complex interactions. Mental health variables are expected to have complex dependencies with potential confounding factors, mediation, and intercausal dependency; therefore, we extended our association analysis with data-driven structure learning of a BN. We preferred this approach over black box machine learning and standard statistical modeling for several reasons. Structure learning allows us to discover and model confounding factors transparently, whereas black box machine learning models such as RFs and gradient boosted machines are not well suited for transparent reasoning. Standard statistical approaches make it humanly impossible to model interactions among hundreds of variables. Structure learning allows discovery and dissection of interactions into mediation, confounding, and intercausal effects. The challenge of incorrect learning is addressed by ensembling many BNs (101 in our case) and choosing the ensemble voted structure. Our artificial intelligence (AI) approach has earlier been validated for public health problems [26,27], and this study demonstrates the underexplored potential of such an approach in complex mental health scenarios.

Our study has a few limitations. Establishing causal inference in cross-sectional data is nearly impossible, and we acknowledge the possibility of confounders. However, this was precisely the reason we chose the structure learning approach, as some of the confounding influences can be transparently discovered and explained. The ensemble voted structure over the sufficiently large number of bootstrapped structures is expected to be robust, as a set of 101 BNs was found to be sufficiently large enough for this study to address the challenge of incorrect learning. Our approach is best suited as a probabilistic reasoning model to explain mental health determinants and to make predictions, a useful outcome in COVID-19–induced mental health morbidity. We could not explain why anxiety levels may be lower in persons with pre-existing cancer or hypertension. This may be a result of reduced work environment–related stress or more contact with family members at home. However, the current data set is not suited to address this at a finer level of explainability. In addition, we could not comment upon the temporality and persistence of these effects. Our results are currently limited to only one geography (ie, the United States). However, the relatively large sample size and multiethnic involvement in the survey makes the model representative for most of the ethnicities and influences across the United States; hence, it is likely to hold true in the United States. Finally, we believe that our study contributes to the use of explainable AI to predict mental health at a population level using survey data, hence making it broadly applicable. Survey data sets are notoriously noisy, and our approach achieved a balance between knowledge discovery and a predictive accuracy of 80%, thus establishing a baseline under a novel scenario. Our algorithms can be used as a screening method for identifying individuals who need help, and further studies with additional measurements and features may increase the accuracy of predictions. Therefore, predictive models for screening and assessing the mental health impact of COVID-19 is a crucial step toward proactive management and prevention of psychiatric comorbidities as populations continue to fight the pandemic.