This study used a mixed methods approach to analyze PPU data from X. The Quantitative Analysis Methods section introduces the quantitative analysis methods applied in this study, whereas the Qualitative Analysis Methods section details the qualitative methods used. For the quantitative analysis, 4848 tweets were analyzed. From this data set, 500 PPU tweets were randomly selected for more in-depth qualitative analysis.

The quantitative analysis leveraged NLP techniques to elucidate the attributes and characteristics of the corpus of tweets. We used 2 key NLP techniques: frequency analysis to understand tweet trends, bigrams, trigrams, and sentiment analysis for the PPU-related sample of 4848 tweets, which we also used to compare their sentiment differences against a control group using a set of random 4848 non-PPU tweets extracted from X during the same period [21,22].

Before applying NLP techniques, proper cleaning was performed on the tweet data set to ensure the accuracy of the NLP outcomes. This data cleaning was solely performed for the quantitative analysis to apply various ML algorithms to analyze our tweets. The process involved tokenizing cleaned tweets into arrays in which each word represented an element. We removed stop words such as a, is, and the, which offer limited insights into tweet sentiment. To avoid the potential loss of sentiment intensity when lemmatizing words, we used the raw text as input for our sentiment analysis. In addition, non–American Standard Code for Information Exchange characters were removed from the text.

Sentiment analysis, which classifies the expressed sentiment in text, was conducted using the Valence Aware Dictionary and Sentiment Reasoner (VADER), which is designed to account for nuances such as punctuation, capitalization, and even emojis, thus providing a more accurate reflection of sentiment in the text. VADER, a lexicon and rule-based tool tailored for social media, assesses the polarity (positive, negative, and neutral) of each tweet and assigns a sentiment score ranging from –1 to 1. For instance, in the tweet “I hate and discourage the use of drugs while pregnant,” it would highlight “hate” and “discourage” as negative, whereas in “I enjoy drugs and alcohol,” it would identify “enjoy” as positive. On the basis of established thresholds, tweets with scores of >0.05 are deemed positive, tweets with scores of <–0.05 are deemed negative, and those in between are deemed neutral [23]. Sentiment analysis has broad applications, including in marketing, politics, and health care, and is particularly effective in X-based health research [21,22]. Thus, in this study, we aimed to apply VADER sentiment analysis to identify the prevalent sentiments associated with the most frequently reoccurring substance combinations in PPU-related tweets.

Furthermore, we conducted frequency analyses, focusing on trend, bigram, and trigram analyses. Bigrams are sets of 2 words together, and trigrams are sets of 3 words together. This approach sheds light on prevalent discussion topics within the data set without necessitating a manual review of each tweet; as noted by Sarker et al [14], such analysis provides meaningful insights into the textual content.

Qualitative analysis enabled a deeper exploration of the underlying semantics of individual tweets. By applying thematic content analysis, we gained deeper insights into PPU-related themes, subthemes, and user behaviors. This method, recently applied in the health care field, helps distill textual data into coherent themes. Thematic analysis emphasizes emergent themes and subjectivity, whereas content analysis focuses on explicit content and objectivity using either inductive or deductive approaches. Inductive content analysis derives concepts, categories, and themes directly from the data, whereas deductive analysis uses predefined frameworks [24,25].

For this analysis, we selected a random sample of 500 PPU tweets from our 4848 PPU-related tweets. In total, 2 independent researchers conducted the qualitative analysis using inductive coding techniques. The tweets were categorized into themes and subthemes using the NVivo software (version 12.0; Lumivero). The researchers started coding tweets together to ensure similar thought processes, followed by independent coding interspersed with periodic meetings to reconcile differences and jointly analyze more complex tweets. After coding all 500 tweets, we assessed interrater reliability using the κ score to determine the degree of agreement between the 2 coders and account for the probability of coding similarly by chance [26]. We computed 2 κ scores using NVivo’s built-in functions: one for the tweets in the main themes and the other for substance combinations. The resulting average κ score was 0.6940 for differentiating tweets about single-substance use, polysubstance use, and non–substance use, achieving reasonable agreement between the 2 coders. Further investigation was conducted to find the average κ score between the combinations of substances. For substance combinations, the average κ score was 0.7748, reflecting a satisfying agreement given that previous studies have identified a κ score of 0.60 as moderate [26] and 0.79 as significant [14].

We established a codebook for this thematic content analysis. Figure 1 illustrates our coding structure with the first few layers of the multilayer codebook. Each tweet was first classified as non–substance use, single-substance use, or polysubstance use. Tweets regarding polysubstance use were further categorized by the specific combination of substances discussed. The final coding structure contains >3 levels or subbranches, so Figure 1 visually represents the general coding process for our qualitative analysis.

To identify specific combinations of PPU, we analyzed commonly mentioned substances from the PPU tweets. We extracted combinations of substances that were both mentioned explicitly in a tweet; for instance, a tweet containing the words “alcohol” and “drug” was categorized into the “alcohol and drugs” combination. We then conducted sentiment analysis on the most frequently reoccurring substance combinations in our data. We further quantified sentiment differences using the chi-square test for the frequency counts of sentiment categories—positive, neutral, or negative—across PPU and non-PPU tweets. This statistical test was crucial for identifying significant variations in sentiment distributions between the 2 groups.

Table 1 displays the overall VADER sentiment analysis results for PPU tweets and random non-PPU tweets. To assess the sentiment distribution differences between PPU and non-PPU tweets, we applied the chi-square test and reported the results in Table 1. Notably, PPU tweets had a significantly lower percentage of positive sentiments (1884/4848, 38.86%) compared to the non-PPU tweet sample (2206/4848, 45.5%), with a P value of <.001, indicating that positive sentiments are less prevalent in PPU tweets. Conversely, negative sentiments were more common in PPU tweets (2490/4848, 51.36%) than in the non-PPU sample (1323/4848, 27.29%), with the statistical significance underscored by a P value of <.001. Neutral sentiments were significantly rarer in PPU tweets (474/4848, 9.78%) as opposed to in the control sample (1319/4848, 27.21%), also with a P value of <.001, suggesting a lower occurrence of neutral sentiments deemed useful for PPU analysis.

Table 2 presents the sentiment analysis for the top 5 most reoccurring substance combinations. Regarding “alcohol and drugs,” “meth and drugs,” “weed and drugs,” and “cocaine and drugs,” the sentiment of over half of these substance pairs was negative, and the sentiment of <40% was positive, except for the sentiment of the “weed and meth” combination, which was mixed (positive: 70/160, 43.7%; negative: 76/160, 47.5%; neutral: 14/160, 8.8%).

Next, we illustrate the overall tweet distribution of our main PPU tweets from May 2019 to October 2021 in Figure 2. We further present the frequency of the most popular bigrams and trigrams in PPU tweets (ie, 2- and 3-word sequences of words, respectively) in Table 3 and 4. Regarding the bigram frequencies listed in Table 3, the 2 apparent themes pertain to drug use and family dynamics. The most common bigrams were “drug alcohol,” “drug use,” and “smoke weed.” These bigrams provided a glimpse of the major substance combinations that we were able to further identify through qualitative analysis, discussed in the next section. In addition, there was a multitude of bigrams alluding to domestic situations (eg, “family member” and “child abuse”). Regarding the trigrams listed in Table 4, the most common ones were related to drug abuse. A large quantity of the results were trigram permutations of the words “drug,” “alcohol,” “abuse,” and “addiction.” Table 4 presents these polysubstances containing the dominant substances.

Overall, among the 4848 PPU tweets, we found a complicated polysubstance use among pregnant individuals. Of the 4848 tweets, 4134 (85.27%) tweets contained 2 substance keywords, 539 (11.12%) included 3 substance keywords, and 175 (3.61%) mentioned >3 substances. Most PPU patterns turned out to be paired with 2-substance use. The top 5 substance combinations were alcohol and drugs, meth and drugs, weed and drugs, cocaine and drugs, and alcohol and meth, listed in Table 2. Table 5 further details the percentages of tweets ranked as the top 11 substance combinations. Figure S1 in Multimedia Appendix 2 visually presents the overall substance combinations in the PPU tweets. Table S1 in Multimedia Appendix 2 further summarizes the top 60 substance combinations included in the “other” category, including other identified double and triple substance combinations and their percentage distributions in this study. The alcohol and drugs combination was the most frequently reoccurring polysubstance group. This further supports the common perceptions about alcohol’s popularity and its widespread acceptance, which has transformed general opinions to no longer consider it as a drug, as indicated in the following tweet quote, also indicating that legal substances tend to be discussed more openly and frequently on social media:

We randomly selected a sample of 500 PPU tweets for a more in-depth thematic analysis. The tweets were initially coded into 3 categories depending on the substances mentioned (ie, nonsubstance, meaning that no substance was mentioned in the tweets; single substance, indicating that only 1 substance was discussed in the tweets; and polysubstance, representing multiple substances mentioned in tweets). PPU tweets were categorized based on the substance name and combination with another substance. We used 3 coding levels to provide an overall visual structure of our codebook illustrated in Figure 1 in the Qualitative Analysis Methods section, and we found more subthemes within each category to allow the researchers to capture the exact essence of each tweet.

Table 6 contains the definitions for each identified theme category with example tweets. Following our coding process (Figure 1), the 2 researchers independently coded and identified themes and subthemes for the sample of 500 PPU-related tweets. There may be some overlaps in the categories; for example, a tweet can start discussing single-substance use and then later discuss polysubstance use. We coded this type of tweets as polysubstance use because they discussed both drug and alcohol use.

In addition to main-level categories or themes, we further coded the contents into subthemes to better illustrate the discussions on X. For the main PPU themes identified, we delved into specific drug combinations observed in the tweets. The most frequently reoccurring combinations were alcohol and drugs, alcohol and cannabis, and alcohol and nicotine. Table 7 presents a brief description of each, agreed upon by the 2 coders, along with example tweets. The κ score for the substance combinations was 0.7748.

To extract more meaningful insights from each thematic area, we created additional subthemes within the identified main substance combinations. The most frequently reoccurring topics were lifestyle, perceptions of others’ use, legal ramifications, and public health. These prominent subthemes were based on the frequency of tweets in each category without necessarily pertaining to PPU. Broad definitions of each of the leading subthemes can be found in Table 8 along with and example tweets.

There are instances in which the subtheme breakdowns contained additional themes within them. The subtheme lifestyle contained more tweets with positive sentiment than negative sentiment (14 vs 12 tweets). The positive subtheme illustrated support for substance use when X users expressed personal experiences and normalcy with substance use. Meanwhile, the negative subtheme discussed how others’ use has affected the tweeters’ lifestyle choices. Lifestyle was the most frequently occurring subtheme within the major substance combinations, and thus, individuals were comfortable discussing their substance use through social media. Many individuals discussed daily substance use, social substance use, and family’s or friends’ substance use as it pertained to lifestyle decisions.

Another common theme throughout perception of others’ use was “use while pregnant.” The total number of tweets in this subtheme of the major substance combinations discussed was 33. Additional themes allow for more accuracy and ensure that the exact topic discussed is being captured, further assisting the discovery of more specific recurring themes. The prevalence of the subsubtheme perceptions of others’ use can be attributed to the intrinsic purpose of social media platforms such as X. Oftentimes, these platforms are used for individuals to voice their opinions of disagreement and concession. Some users stood up for the pregnant women while still holding their stance that substance use should be avoided:

Others chose to adhere to their beliefs and shame individuals who thought otherwise and even went as far as to suggest aborting the pregnancy if it was believed that substances were used while pregnant:

Discussions of abortion and miscarriage laws were the main topic behind the subtheme of legal ramifications. Criminal charges can be brought against a mother who uses substances while pregnant; some of those charges include child endangerment, attempted aggravated child abuse, child neglect, and even manslaughter [12]. Many of the tweets that fell under this subtheme were related to when a woman could be charged for a miscarriage if she ingested substances during her pregnancy. One discussion revolved around whether a woman would be held responsible for a miscarriage due to substance use if she was unaware of her pregnancy.

Users who are currently pregnant and struggling with substance abuse may find all the negativity and judgment discouraging, thus making them hesitant to reach out for help due to fear of persecution. Researchers who studied secrecy versus disclosure found several explanations for apprehension in pregnant women with substance use issues, specifically shame and fear of judgment as motives for not reaching out to family and physicians [11]. Mothers who use substances are ashamed of the harm they are potentially causing their children while they live in fear of criminal charges they could be subject to, along with the possibility of having their children taken away from them [12,27].

The following is one of many tweets discovered in our corpus that project negative opinions of women who partake in substance use during their pregnancy. It exemplifies the stigmas that surround the topic:

The quantitative sentiment analysis revealed clear variations in user sentiments toward PPU. Tweets related to PPU presented a significantly lower percentage of positive sentiments compared to those unrelated to PPU. In addition, PPU-related tweets presented a significantly higher percentage of negative sentiments. Furthermore, neutral sentiments were significantly less frequent in PPU tweets—only 9.78% (474/4848) compared to 27.21% (1319/4848) in the non-PPU sample—highlighting a lower tendency to express neutral views on PPU. We also identified a set of complicated substance use combinations in the 4848 PPU tweets, among which paired substance use (4134/4848, 85.27%) dominated, with “alcohol and drugs,” “meth and drugs,” “weed and meth,” and “cocaine and drugs” as the top 5 paired substance types. The combination of “alcohol and drugs” use was ranked as the most frequently reoccurring combination in PPU discussions on X. A total of 11.12% (539/4848) of the PPU tweets mentioned 3 substances, and 3.61% (175/4848) included >3 substances.

The results from thematic content analysis illustrated that alcohol and drugs were the driving forces of the substance combinations. These substances were the most frequent in the combinations discovered through the qualitative analysis. Alcohol is a widely accepted substance due to its use in social gatherings and personal consumption as a legal substance if users’ ages meet legal requirements, thus making it a leading substance in this study.

On the basis of the qualitative sentiment analysis, PPU had an overall negative sentiment in tweets, implying that substance use during pregnancy was mostly not supported in the gathered tweets. Except in specific situations, X users essentially perceived that PPU is dangerous and should be avoided. Some users even suggested abortion to avoid adverse outcomes for the infant. Consistent themes from this intersection include how others viewed the pregnant individual’s PPU regarding the negative health outcomes for the child when substances are used during pregnancy and the significant stigma related to using substances while pregnant. The subtheme public health provided some knowledge to educate pregnant individuals on avoiding, reducing, and intervening in PPU and offer support to this vulnerable population.

Furthermore, a high volume of tweets pertaining to legal ramifications (n=28) reflected impassioned speakers on social media platforms. Users will tweet their opinions to corroborate, advertise, and support their views on the current state of US politics and policies. One major topic dealt with miscarriage charges for pregnant individuals who used substances during the pregnancy and how their behaviors may have led to the miscarriage. This can be connected to the fear of disclosing their substance use patterns while pregnant.