Machine learning sheds light on mental health challenges faced by HCWs during COVID-19 pandemic

In a study published in JMIR AI, researchers assessed anxiety and depression faced by healthcare workers (HCWs) in the United States during the coronavirus disease 2019 (COVID-19) pandemic.

Using machine learning methods, they showed how the issues that healthcare workers experience are unique and highlighted ways to support this indispensable workforce more effectively.

 

Background

Medical professionals are more vulnerable than the general population to mental health challenges such as depression, anxiety, and suicidal ideation. COVID-19 increased the stress and workload faced by HCWs further. As the pandemic surged, the number of patients exceeded available beds, and hospitals were forced to operate over capacity.

HCWs worked longer hours under adverse conditions, including equipment and resource shortages, which forced them to ration care and make difficult decisions.

As frontline workers, they were more exposed to the virus and often had limited access to masks and other protection. Like many others, they also lost the support of social and familial networks due to strict quarantine guidelines.

HCWs suffering from depression and anxiety are more likely to commit errors, inadvertently jeopardizing patient safety. Improving their well-being is vital to strengthening the healthcare system as a whole.

This calls for more research in order to gain a thorough understanding of the mental health challenges HCWs face and provide them with the support that they need. Such interventions will be instrumental in making the health system resilient to future pandemics and other disruptions.

About the study

Researchers obtained the treatment transcripts of 820 HCWs who received digital psychotherapy from licensed providers from March to July 2020. These transcripts were de-identified to protect patient privacy.

HCWs included providers such as physicians, residents, nurses, social workers, and emergency medical service providers. They were all self-referred and had active National Provider Identifiers (NPIs).

They received therapy through an initiative to provide free treatment to HCWs for one month. The telehealth platform that donated these services also treats non-HCWs.

To identify how the challenges faced by HCWs differed from those of the general population, researchers matched each provider to a non-HCW based on similarities in symptoms, demographics, treatment start date, and state of residence. The non-HCWs included in the study were English-speaking US residents with access to the Internet.

Before they received therapy, all patients were assessed for depression and anxiety by a licensed provider. The Patient Health Questionnaire-9 was used to measure depression symptoms, and a General Anxiety Disorder Scale-7 evaluated symptoms of anxiety.  

They were excluded from the study if they (1) required hospitalization, (2) were having suicidal thoughts, or (3) were experiencing bipolar, substance abuse, and other disorders.

Researchers used a heuristic classification algorithm to obtain each HCW’s profession from the transcript. They further processed the de-identified transcripts by converting words to their root forms to create a ‘vocabulary.’ They removed empty transcripts and words that emerged from fewer than 50 documents.

This resulted in 1,208 terms from the 820 HCW transcripts and 1,259 from the 820 non-HCW transcripts. Structural topic modeling (STM) methods were then used to identify topics raised by patients and the associations between topics and levels of depression and anxiety.

Results

HCWs were predominantly female (91%) and aged, on average, 31.3 years old. New York State and California accounted for more than one-quarter of the sample. Slightly over half of the HCWs were nurses, while less than 20% were physicians.

Notably, 35.2% of HCWs reported that this was their first experience with psychotherapy. Slightly over 56% of the HCW patients were diagnosed with anxiety disorders, while only 8.2% were diagnosed with depressive disorders. Prior to treatment, 601 out of 820 HCWs (73.3%) had either depression or anxiety.

STMs showed that HCWs frequently brought up four topics related to healthcare provision. The topics exclusively mentioned by them included (1) fears related to the coronavirus, (2) their work in intensive care units (ICUs) and hospital floors, (3) masking and patients, and (4) their roles (such as attending or resident).

In stark contrast, the non-HCWs only mentioned one topic related to pandemic anxiety and one topic related to their employers.

With regard to mental health, both HCWs and their matched controls brought up five topics, discussing panic attacks, disturbances to their moods, and experiences of grief. HCWs also frequently mentioned disruptions to their sleep.

Providers experiencing moderate to severe depression or anxiety were more likely to discuss the hospital or areas such as the ICU. Compared to matched controls, HCWs were also more likely to mention mood alterations or sleep disruptions.

Conclusions

By comparing 820 healthcare providers with 820 matched non-HCW patients receiving therapy from the same platform, researchers of this study used machine-learning computational linguistics methods to show that HCW patients showed unique associations between psychiatric symptoms and their work. The findings show that the pandemic increased work-related stress levels that HCWs routinely face, highlighting the need to prioritize their mental health.

The authors acknowledged the study’s limitations, identifying avenues for future research. Since patients were self-referred, the researchers were unable to include those with limited access to virtual therapy.

There was a clear skew in the sample towards female providers, particularly nurses, indicating the need to reach more physicians and male practitioners. Further studies could also include more complex linguistic models and allow the analysis of non-English transcripts.

Despite these shortcomings, it is clear that the study provides actionable evidence of the unique challenges faced by HCWs during the COVID-19 pandemic.

It also demonstrates how machine-learning algorithms can be used to process and analyze large datasets and guide clinical interventions while preserving the privacy of study participants.

Written by

Priyanjana Pramanik

Priyanjana Pramanik is a writer based in Kolkata, India, with an academic background in Wildlife Biology and economics. She has experience in teaching, science writing, and mangrove ecology. Priyanjana holds Masters in Wildlife Biology and Conservation (National Centre of Biological Sciences, 2022) and Economics (Tufts University, 2018). In between master's degrees, she was a researcher in the field of public health policy, focusing on improving maternal and child health outcomes in South Asia. She is passionate about science communication and enabling biodiversity to thrive alongside people. The fieldwork for her second master's was in the mangrove forests of Eastern India, where she studied the complex relationships between humans, mangrove fauna, and seedling growth.