Since late 2019, the ongoing COVID-19 pandemic has proved to be extremely disruptive; the pandemic has resulted in high morbidity and mortality rates, as well as economic and mental health consequences. These issues pose a challenge for mental health services, as little is known about SARS-CoV-2 and the psychological impact that the COVID-19 crisis has on medical professionals, essential workers, unemployed individuals, and other people who engage in physical distancing. Early reports have suggested that mental health professionals face considerable challenges due to the lack of information and established guidelines for assessing and treating patients with COVID-19 [1]. Psychological sequelae to major events like the COVID-19 pandemic vary greatly [2-4]. This makes it difficult to assess the full range of symptoms that are potentially related to the COVID-19 crisis, and to track the course of COVID-19–related reactions over time. The use of comprehensive symptom checklists and exhaustive clinical interviews poses a considerable burden to clinicians and respondents alike. However, this burden can be avoided if the most common symptoms are known ahead of time.

A major obstacle to identifying the most relevant symptoms for screening is the amount of time it takes to amass clinical observations for a suitably large patient population. Large patient populations are needed for determining the full scope of patients’ reactions. It is difficult to differentiate peripheral symptoms from central and pathogenic symptoms [5]. It is also difficult to differentiate symptoms that are ordinarily reported by a diverse, treatment-seeking population from symptoms that are more closely associated with COVID-19–related concerns. However, there is a lack of this type of data for the ongoing COVID-19 crisis. Data that are generated by large telemental health services that remotely deliver care (ie, US and global telemental health services) may be helpful in describing the full complexity of patients’ clinical presentations. Text messages between patients and therapists (ie, text messages that are a part of intake and treatment procedures) offer the most useful data. These data are relatively unstructured compared to standardized symptom measures, but they offer the advantage of capturing patients’ experiences more comprehensively. This allows therapists to differentiate symptoms that are reported in conjunction with mentions of COVID-19, from symptoms that are reported by individuals who seek care for other reasons. Natural language processing (NLP) methods refer to a broad set of methods that have been designed to analyze unstructured textual data. These methods range from simple methods that search for specific words in a block of text, to more complex neural network models that extract the meaning of certain statements by analyzing the larger context of a text corpus.

The aim of this study was to identify major themes and symptom clusters in the text messages that patients send to therapists. We assessed patients who were seeking treatment for pandemic-related distress on Talkspace, which is a popular telemental health platform. To achieve our objective, we differentiated symptoms that were associated with the pandemic from symptoms that were only experienced by individuals who seek treatment, by investigating the relationships among words that were associated with mentions of COVID-19. This is impossible to do when only relying on structured symptom measures that do not specify whether symptoms relate to the pandemic or some other cause. Therefore, we used a multistep process that involved brief symptom measures to determine the relationship between the pandemic and common, self-reported anxiety and depression symptoms; identify patients with COVID-19-related concerns; and isolate words that highly correlate with mentions of SARS-CoV-2. We categorized these words by using a digital phenotyping process for determining the prevalence of clinical and nonclinical themes, to ultimately identify symptoms in diagnostic categories that are not reflected in structured measures for common anxiety and depression symptoms. Our study demonstrates that our method has strong face validity and high levels of clinical interpretability. Therefore, our method can potentially be used to inform decisions on structured measures for tracking responses to the pandemic over time.

In this study, we investigated the relationship between the COVID-19 pandemic and the intake anxiety and depression symptoms of treatment-seeking patients on a digital mental health platform. We identified a significant and noticeable increase in anxiety symptom severity, but not in depression symptom severity. We also applied machine learning methods to a large body of treatment transcripts via NLP methods, to identify additional symptoms that were associated with mentions of COVID-19, but would have been missed by symptom measures that assess anxiety or depression alone. These additional symptoms included those that were associated with other diagnostic categories, such as acute stress, posttraumatic stress disorder, grief, obsession-compulsion disorder, insomnia, hypochondriasis, nonsuicidal self-injury, and paranoia. In some ways, a more complex symptom profile can be generated via dimensional approaches to psychopathologic nosology [20,21], which focuses more on symptoms and functions rather than diagnostic categories. Our study suggests that tracking the lasting psychological impact of COVID-19 requires measures for a variety of symptoms from several disorders. To date, survey-based studies have accounted for a mix of depression, anxiety, insomnia, and stress-related conditions [22-24]. However, these studies have not reported data on the other symptoms that we identified in this study. Constructing an appropriate measure—whether by combining self-reported ratings with clinicians’ ratings or ratings from other sources, using the advantages of ecological momentary assessments, or developing other strategies for mitigating recall bias—is beyond the scope of this study. However, our study highlights that constructing appropriate measures is an important next step in applying our findings to practice.

An assessment that is composed of symptom questions that are informed by the appropriate measures could ultimately identify subpopulations of patients with different symptom profiles. This would assist with individualizing treatments and tracking heterogeneous responses to clinical interventions. For example, social isolation and loneliness are distinct risk pathways for suicide. Therefore, these risk pathways should be assessed as distinct behaviors, to inform treatment planning [25]. This is also the case for individuals with comorbid psychiatric disorders. Comorbidity is common, yet without pandemic-specific, longitudinal assessments, true comorbidity could be conflated with changing symptom constellations for the same underlying pathology. This has been exemplified in cases that demonstrate the dynamic interplay of bipolar disorder and anxiety symptoms [26]. Selecting symptom questions based on the rapid digital phenotyping methods that we implemented in this study can help reduce the burden on respondents, and provide a broader dimensional approach to monitoring psychopathology (ie, an approach that focuses on symptoms rather than diagnostic categories) [20]. Indeed, one of the goals of the National Institute of Mental Health has been to analyze disorders via a dimensional approach that does not rely on disorder categories, but instead draws on big data (ie, large clinical datasets) to increase our understanding of the underlying mechanisms of health and illness [27,28]. The data reported in this study are an important first step in the efforts for understanding symptom clusters that are associated with the pandemic, guiding the discovery of pathogenic mechanisms, and informing personalized interventions that maximize treatment benefits [29-31]. A critical next step for research is continuing to evaluate COVID-19 symptoms after vaccination and other programs begin to lower SARS-CoV-2 infection rates and reduce the threat of COVID-19. Commonly identified behaviors, such as social withdrawal, extreme anger, and COVID-19–related paranoia, may diminish after the pandemic and reflect adaptive responses to the pandemic. If, on the other hand, these symptoms persist, the possibility that other pathology mechanisms are at play increases, and further research would be warranted.

In our study, the lack of terms with the word stem “lone” may be in line with studies that have reported that loneliness is unlikely to be a major factor in COVID-19 pandemic–related psychological distress [24]. The lack of these terms also suggest that people are in fact isolated from others, given the few mentions of social topics in the text messages of patients with COVID-19–related complaints. The patients in this study may not yet think that isolation is similar to loneliness. The 18.5% (40,448/219,156) of patients who mentioned the pandemic exhibited a substantial increase in disease burden over and above that of the prestudy patient population that was already undergoing treatment. This finding corresponds with the increased number of new COVID-19 cases that was reported on the Talkspace platform. An important feature of this study is that we distinguish and quantify patients who seek care for COVID-19–related concerns, instead of patients who would have sought care without the influence of pandemic-related stressors. In this study, although anxiety symptom severity started to return to pre–COVID-19 pandemic levels between May 5, 2020 and May 30, 2020, we observed the opposite trend on early June 2020. It is thus advisable to continue focusing on patients who experience psychological symptoms and require treatment as a result of the pandemic.

Although this study offers a novel method and dataset for rapidly phenotyping COVID-19–related symptoms, it is not without limitations. First, our results may not be generalizable beyond the population of individuals who seek treatment through digital platforms. Second, our analyses relied on the longitudinal data of a convenience sample that self-reported their symptoms. We did not assess referral sources or use random assignment methods. Of particular note is the large number of women in our sample. However, this is consistent with existing data on the use of telemedicine services for routine care [32]. Despite these limitations, our results demonstrate the utility of large, unstructured data in rapid digital phenotyping methods for identifying psychological symptoms that are associated with patients’ COVID-19–related concerns, but are missed by standard depression and anxiety screening methods. The symptoms we identified in this study can be used to inform standard symptom surveys. Our study demonstrates how digital phenotyping can assist in and accelerate the development of traditional monitoring tools that do not require the use of digital therapy platforms or large amounts of textual material, avoid the potential for unwanted monitoring among technology users, and ensure that monitoring is an overt process (ie, people who are monitored are aware of being monitored).