The Emotional Landscape of Pregnancy and Postpartum during the COVID-19 Pandemic in Italy: A Mixed-Method Analysis Using Artificial Intelligence

Italy, as the first Western country severely affected by COVID-19, experienced widespread uncertainty, fear, and significant mental health burden. Prenatal and postnatal care were disrupted, and pregnant and postpartum women faced elevated risks for adverse obstetric outcomes and mental health problems (depression, anxiety, PTSD). Prior work by the authors during Italy’s first lockdown showed substantial psychological distress among pregnant women and suggested exclusive breastfeeding as a protective factor. The present study aims to provide a more comprehensive, nuanced understanding of the emotional experiences and coping strategies of pregnant and postpartum women in Italy during the first pandemic months, focusing on emotions surrounding birth and the influence of communication from healthcare professionals (HCPs), media, and peers.

The paper situates the study within evidence showing increased psychological distress in the general population during COVID-19 and specific vulnerability among perinatal women, including higher risks for anxiety, depression, PTSD, and adverse maternal-neonatal outcomes. Prior Italian studies indicated heightened perinatal distress during lockdown and subsequent periods; social restrictions increased depression risk in low-risk pregnancies, while hospitalization in high-risk pregnancies sometimes provided social support. Literature also links infodemics and media exposure to fear and confusion, highlights the importance of HCP communication quality, and notes the role of coping strategies, social support, and dispositional optimism in mitigating distress.

Design: Secondary analysis of COVID-ASSESS, a national cross-sectional survey administered online during the first COVID-19 wave in Italy (March–May 2020) via the Ciao-Lapo network. Participation was voluntary and self-selected with online informed consent. Original sample: 2448 women (1307 pregnant, 1141 postpartum). Eligibility for this post hoc analysis required responding to at least one qualitative question. Final analyzed sample: 1774 women (1136 pregnant, 638 postpartum). Ethics approval: University of Florence Ethics Committee (Prot. 0064823, n. 61, 82, 03.20). Data are available in Mendeley. Survey content included socio-demographics, obstetric history (including losses), psychopathological history, birth expectations pre/post COVID-19, concerns, postpartum health/infant feeding, perceptions of media/HCP communication, and psychometric evaluations. Emotion classification: Zero-shot prompt-based classification using GPT-3.5-turbo (via GPT for Sheets) for eight emotions (anger, anticipation, joy, trust, fear, surprise, sadness, disgust). Preprocessing removed irrelevant text, lowercased, tokenized, and removed stopwords/punctuation. The model outputted binary presence/absence strings for each emotion for pre-COVID and post-COVID contexts across 1774 rows; outputs were imported into StataBE 18 and converted into emotion variables. Sensitivity analysis: A small random subsample (n=20) compared GPT-3.5-turbo to GPT-4, showing 99.4% concordance; GPT-3.5 was faster and cheaper, so it was used for the full analysis. Thematic analysis: GPT-4 was used to identify and score themes from blocks of qualitative text (healthcare communication: 5434 words; media: 3828 words; peers: 3767 words; open-ended responses: 46,542 words). Steps included tokenization, frequency analysis, semantic grouping into themes, scoring theme prominence on a 0–10 scale based on frequency proportions, and extracting representative quotations (checked and translated by authors). A thematic map was created using Xmind based on AI-identified connections, literature, and researcher experience. Quantitative sample descriptors included age, education, trimester/postpartum infant age, parity, prior loss, and psychopathological history; group comparisons between pregnant and postpartum women were performed (no significant baseline differences reported).

Sample: 1774 women; 1136 pregnant (64.0%), 638 postpartum (36.0%); mean age 33.6 years (SD 4.8). Education was generally high; 64.2% multigravidae; 38.0% prior pregnancy loss; 44.5% reported psychopathological history. No significant baseline differences between pregnant and postpartum groups. Emotions related to birth (before vs during pandemic): Before COVID-19, most common emotions were trust (n=875), anticipation (n=735), and joy (n=703). During COVID-19, most common were sadness (n=760), fear (n=696), and anticipation (n=449). Largest changes: trust decreased by 49.3%, joy decreased by 36.4%; sadness increased by 52.3%, fear increased by 49.3%. Thematic findings (scores 0–10 indicate prominence): - HCPs and communication: Fear/anxiety (8/10); uncertainty/confusion (7/10); emotional support/reassurance (6/10); professionalism/competence (5/10); distance/detachment (4/10). - Media communication: Fear/anxiety (10/10); confusion/contradictions (8/10); sensationalism/alarmism (7/10); misinformation/inaccuracy (6/10). - Peers communication: Remote communication (8.5/10); adaptation and coping strategies (7.5/10); emotional support (7/10); information sharing (6/10); adaptation to the new normal (5.5/10); fear/anxiety (5.5/10). - Open-ended responses: Fear and anxiety (9/10); loneliness and isolation (8/10); gratitude and hope (6/10); disappointment and sadness (5/10). Word clouds and supplemental table S1 illustrated most frequent words pre- and during-pandemic. Overall, the pandemic shifted the emotional landscape from predominantly positive (trust, joy, anticipation) to negative (sadness, fear), with communication quality and content across HCPs, media, and peers shaping experiences.

The study confirms and extends earlier preliminary findings by the authors, demonstrating a marked shift from joy and trust before the pandemic to fear and sadness during the acute phase. These changes align with elevated maternal and neonatal risks associated with SARS-CoV-2 infection and literature linking negative or unexpected childbirth experiences to distress. Communication quality profoundly influenced emotions: unclear, contradictory, or rapidly changing guidance from HCPs fostered uncertainty; media infodemics amplified fear through sensationalism and misinformation; peer networks mitigated isolation through remote contact and emotional support. Hope and optimism were fostered by empathetic, competent HCP communication and supportive social ties. Loneliness and isolation were common due to restrictions and care modifications, potentially increasing mental health risks. Coping strategies included seeking reassurance from HCPs, limiting distressing media exposure, focusing on positive news, and leveraging virtual social connections. The findings highlight the need for clear, consistent, empathetic information, robust psychosocial support during crises, and attention to expectation–experience mismatches around pregnancy and birth.

The pandemic significantly altered the emotional landscape of pregnant and postpartum women in Italy, with fear and anxiety becoming dominant and displacing pre-pandemic joy and trust. Communication played a dual role: media-driven misinformation and alarmism exacerbated distress, while supportive, clear interactions with HCPs, family, and peers helped buffer negative emotions. Targeted mental health support, improved crisis communication strategies, and accessible social support systems are essential for protecting perinatal well-being in future public health emergencies. Future research should refine AI-assisted qualitative methods, evaluate interventions to enhance optimism and coping, and develop evidence-based communication frameworks to reduce uncertainty and fear among perinatal populations.

AI-based zero-shot emotion and theme classification may be less accurate than fine-tuned or task-specific models and can lack interpretability and sensitivity to contextual nuance. Automated coding may miss subtle meanings or novel insights compared with human qualitative analysis. The study relied on a self-selected online sample during a specific early pandemic window, which may limit generalizability. Although a small sensitivity check showed high concordance between GPT-3.5 and GPT-4, broader validation against human-coded gold standards was not performed.