A separate survey conducted by the Harvard T.H. Chan School of Public Health found pregnant women are suffering from fears of becoming ill with COVID-19, unemployment and financial stress, and pregnancy conditions not being diagnosed in a timely manner due to reduced access to doctors appointments or other healthcare services.

Whilst these surveys provide a snapshot into the broader trends of how the pandemic is adversely affecting pregnant women, the pandemic and its implications are constantly evolving and machine learning offers a more effective way for tracking these changes in real time.

By using machine learning techniques to analyze hundreds of thousands of public social media posts our team at Velmio was able to identify ways in which the events of 2020 have impacted pregnant women. We continuously update this model to ensure that the Velmio app is always addressing the most critical needs of pregnant women today. In this article we share some of our findings that can help health professionals identify key challenges for pregnant women during the pandemic.

Our dataset consists of a few hundred thousand public social media posts downloaded from online communities popular with pregnant women, such as Reddit and dedicated pregnancy forum sites. Over 90% of pregnant women search the internet for health information and a large proportion of these use social media throughout their pregnancy to share their experiences and seek advice and/or emotional support from their peers.

After preprocessing our data, we used a statistical algorithm to identify the key topics discussed by pregnant women. These methods are discussed in more detail at the end of this article, for those interested. We also developed a metric for estimating the popularity of a topic by weighting the number of posts associated with a given topic in proportion with the number of comments (this approximates the level of engagement with the topic). The graphic below shows the average popularity/engagement score for each week of pregnancy for a few of the key topics our model identified.

What's interesting is how topic popularity changes throughout pregnancy. For example, pregnancy tests is the most popular topic of discussion in the first weeks of pregnancy, but becomes less relevant as the positive pregnancy test results arrive (after around week 6). Topics like labor and sleeping difficulties show the opposite trend, becoming more significant in the later stages of pregnancy (or after birth, in the case of sleeping difficulties!). The topic ultrasound tests predictably peaks around the weeks at which pregnant women are likely to get these scans done (most women have 2 ultrasounds during pregnancy, with some receiving a third based on medical needs). Similarly, gender reveal peaks at the stage of pregnancy parents-to-be usually learn the gender of their baby. Meanwhile, nutrition shows a downward trend throughout pregnancy (perhaps as the pregnancy diet becomes routine there's less need to discuss it), but trends upwards after birth (as dietary needs change yet again and women seek guidance on this new phase of their pregnancy journey).

The fact that our model is able to encode these patterns indicated to us that it's working in the way we expect. Furthermore, our model was also able to identify key sentiments expressed in the texts. In the context of mental health, negative sentiment analysis helps us identify texts with anxious or depressive sentiment. One topic in our model stood out as representative of negative sentiment, capturing posts containing words such as "anxiety", "struggle", "stress", "worry", "panic", "scared" and "hard". Interestingly, when plotted against week of pregnancy this topic showed a mostly white noise (random) distribution, either side of week 40 (when childbirth is celebrated and sentiment is predominantly positive). What this means it that anxiety and depression is present at every stage of pregnancy.

With postnatal depression affecting more than 1 in every 10 women, a better understanding of these negative sentiments can aid health professionals and our team at Velmio in developing better programs for preventing and treating depression that occurs during and after pregnancy.

The graphic shown at the beginning of this article shows the subtopics our model identified within the topic of texts with anxious/depressive sentiment, shedding light on the issues that cause considerable anxiety for pregnant women. A key question we explored with our model is how the Covid-19 pandemic has affected the mental health of pregnant women. To conduct this analysis we constructed a dataset of 2019-20 pre-pandemic and 2020 during pandemic posts (i.e. before/after March 2020, when WHO declared Covid-19 a pandemic), applying random sampling techniques to reduce bias from seasonality influences in the stages of pregnancy represented across both datasets.

According to our model, the proportion of posts about anxiety related to health complications has increased by 12% during the pandemic, with the amount of fear of miscarriage posts increasing by 29%, hypertensive disorders by 20%, and anxiety about stillbirth and birth defects by 16%. Posts about lack of family support have increased by 9% and posts related to financial stress have increased by 3.5%. There was a 5.4% increase in posts expressing anxiety about medical test results and a 3.8% increase in posts explicitly mentioning panic attacks. Many of these social media posts identified by our modelling spoke of challenges caused directly by the pandemic:

"I also saw a post on FB last night about a mommy who had to deliver the baby alone since the crisis is so bad and they are not allowing ANYONE in the room! That terrifies me since this is my first baby."

"Anyone else not have their own mama to talk to or share this with? My mom has early onset dementia and is on hospice. I was able to (against all COVID odds) go into her memory care unit yesterday to tell her I was pregnant. Hubs had to be on FaceTime and I was in a mask."

"Not to mention my work just slashed everyone's salaries by 8% yesterday as a result of bad sales because of Covid which doesn't help at all. I'm happy to still be employed but every penny counts and this is just a terrible time for that to happen."

Through our modelling it became apparent that the topics identified could be grouped into several key clusters. This has helped our team identify several focus areas where Velmio's digital health technologies can better support women during pregnancy and reduce anxiety:

Symptoms and medications (e.g. topics identified by our model such as (physical) pain symptoms, severe morning sickness, "normal" symptoms and safe medications): Through our Health Fabric technology, the Velmio app already helps pregnant women track their symptoms and find connections with lifestyle and environmental factors. This helps Velmio users prevent and alleviate their symptoms by discovering approaches that work individually for them, backed by clinical guidelines. As part of our commitment to only ever use data for human good, data from the Velmio app can serve in a large-scale study of pregnancy health. For example, our technology will help in answering the ubiquitous question "is this symptom normal?". A great deal of anxiety is caused by open questions in pregnancy health. We believe that when science finds factual answers this anxiety is reduced because health professionals can deliver confident reassurance to their patients.

Finding support you need (e.g. topics identified by our model such as relationship troubles, societal expectations, lack of family support, financial stress): Digital tools provide communication channels for connecting with health professionals, as well as family, friends, and other pregnant women going through similar experiences. The data analysed in this article is itself enabled by digital communications platforms, such as online forums and social media. The Covid-19 pandemic has seen a surging interest in telemedicine and remote health monitoring solutions, and as a digital platform the Velmio app continues to remove barriers to accessing quality healthcare services.

Pregnancy conditions and complications (e.g. topics identified by our model such as health complications, gestational diabetes, hypertensive disorders and fear of miscarriage): Through advanced lifestyle monitoring, Velmio is the first pregnancy app leveraging the latest innovations in digital health technologies and artificial intelligence to reduce the risk of pregnancy complications. Over time, with more users our technology becomes more sophisticated to offer the first preventative monitoring system that can passively detect a potential health complication from just your smartwatch device. Alongside this, we currently provide tools inside the Velmio app to help women navigating pregnancy with conditions such as gestational diabetes.

How is this analysis done?

The techniques we used here are part of a sub-branch of artificial intelligence (AI) called natural language processing (NLP), which focuses on ways computers can efficiently analyze texts written by humans. We downloaded social media posts from popular web-based pregnancy forums and applied a variety of pre-processing techniques to create a dataset suitable for our analysis. This involves running programable code to convert human language to a format that computers can understand.

We then used a variety of mathematical algorithms for clustering and topic modelling to obtain the results that we report in this article. These algorithms identify statistical trends in the data and group together social media posts that are centered around similar topics, allowing us to study the popularity of certain topics over time. These techniques are known as "unsupervised" learning because the machine discovers patterns and information in the dataset.

The advantage of using machine learning with social media posts is that we can update our model in real-time to always have an accurate view of how the pandemic is affecting pregnant women. Furthermore, the sheer volume of activity that occurs via social media channels provides a large and varied dataset for analysis.

As with any machine learning application, it is important to recognize potential biases and limitations of the analysis. While online forums provide a platform for users around the world to anonymously discuss various topics, it is important to keep in mind that not every pregnant woman uses these forums and so the dataset may not be exactly representative of the pregnant population. Detailed demographic information is not available for the public datasets we used, but since we restricted our analysis to texts written in the English language it is likely our dataset suffers from overrepresentation of users from English-speaking nations and other potential biases. Nevertheless, despite potential limitations, we believe our work provides some important insights that are incredibly useful in understanding the needs of pregnant women today.