Identifying Women with Post-Delivery Posttraumatic Stress Disorder using Natural Language Processing of Personal Childbirth Narratives

Structured Abstract

Background:

Maternal mental disorders are considered a leading complication of childbirth and a common contributor to maternal death. In addition to undermining maternal welfare, untreated postpartum psychopathology can result in child emotional and physical neglect, and associated significant pediatric health costs. Some women may experience a traumatic childbirth and develop posttraumatic stress disorder (PTSD) symptoms following delivery (CB-PTSD). Although women are routinely screened for postpartum depression in the U.S., there is no recommended protocol to inform the identification of women who are likely to experience CB-PTSD. Advancements in computational methods of free text has shown promise in informing diagnosis of psychiatric conditions. Although the language in narratives of stressful events has been associated with post-trauma outcomes, whether the narratives of childbirth processed via machine learning can be useful for CB-PTSD screening is unknown.

Objective:

This study examined the utility of written narrative accounts of personal childbirth experience for the identification of women with CB-PTSD. To this end, we developed a model based on natural language processing (NLP) and machine learning (ML) algorithms to identify CB-PTSD via classification of birth narratives.

Study Design:

A total of 1,127 eligible postpartum women who enrolled in a study survey during the COVID-19 era provided short written childbirth narrative accounts in which they were instructed to focus on the most distressing aspects of their childbirth experience. They also completed a PTSD symptom screen to determine CB-PTSD. After exclusion criteria were applied, data from 995 participants were analyzed. An ML-based Sentence-Transformer NLP model was used to represent narratives as vectors that served as inputs for a neural network ML model developed in this study to identify participants with CB-PTSD.

Results:

The ML model derived from NLP of childbirth narratives achieved good performance: AUC 0.75, F1-score 0.76, sensitivity 0.8, and specificity 0.70. Moreover, women with CB-PTSD generated longer narratives (t-test: t=2.30, p=0.02) and used more negative emotional expressions (Wilcoxon test: ‘sadness’: W=31,017, p=8.90e-⁰⁴; ‘anger’: W=35,005.50, p=1.32e-⁰²) and death-related words (Wilcoxon test: W=34,538, p=3.48e-⁰⁵) in describing their childbirth experience than those with no CB-PTSD.

Conclusions:

This study provides proof of concept that personal childbirth narrative accounts generated in the early postpartum period and analyzed via advanced computational methods can detect with relatively high accuracy women who are likely to endorse CB-PTSD and those at low risk. This suggests that birth narratives could be promising for informing low-cost, non-invasive tools for maternal mental health screening, and more research that utilizes ML to predict early signs of maternal psychiatric morbidity is warranted.

Condensation:

Women with PTSD following childbirth may be accurately identified using natural language processing and machine learning applied to unstructured personal childbirth narratives.

Introduction

Approximately 140 million women give birth every year worldwide, and among them, an estimated one-third experience a highly stressful, potentially traumatic birth.^1–4 The emotional toll of this exposure to trauma can result in a mental illness formally recognized as posttraumatic stress disorder (PTSD) that has been traditionally associated with war, combat, and serious sexual assault.⁵ However, the possibility that childbirth-related trauma could be significant enough to cause PTSD symptoms in postpartum women is lately receiving growing scientific and clinical recognition.^6–8

Of the general population of women giving birth worldwide, ~6% are estimated to experience full childbirth-related PTSD (CB-PTSD).^7,8 This translates to ~8 million women affected in 2022. Women at heightened risk are those with medically complicated deliveries, such as in cases of unscheduled/emergency Cesarean section,^{1, 9} obstetrical complications,⁶ and maternal near-miss.^10,11 Racial and ethnic disparities in experiences of childbirth trauma have also been documented;¹² Black and Latinx women are nearly 3 times more likely to endorse acute stress response to childbirth.¹²Altogether, ~20% of high-risk individuals are likely to endorse CB-PTSD.^8,13

CB-PTSD develops in temporal proximity to the birth of the child, and the child may become a traumatic reminder to the mother. A core complication is impairment in mother-infant bonding. Maternal attachment problems occur across the first postpartum year^14,15 as well as reduced exclusive breastfeeding during an important time for child development.¹⁶ This suggests that CB-PTSD can impede early child development and result in significant public health costs.

An essential step in facilitating maternal psychological adjustment following traumatic deliveries involves early and accurate identification of women with probable CB-PTSD. Accurate screening could be the first step in optimizing opportunities for effective interventions and allocating appropriate resources to targeted, at-risk individuals.¹⁷ The challenges of postpartum mental health screening involve, in part, women’s tendency to under-report their symptoms.¹⁸ Concerns of shame, stigma, and forced separation from infants, as well as poor awareness, hinder screening.^19–21 This suggests that assessment of CB-PTSD derived primarily from patients’ reporting of symptoms could be limited by minimal introspective ability and desirability bias in reporting.

Natural language derived from spontaneous word usage could serve as a marker of well-being and psychopathology.²² In the context of trauma, the personal memory of the event is a central contributor to PTSD etiology and maintenance.^23,24 Accordingly, studies have demonstrated that how individuals recall and narrate traumatic events, including the narrative language, relates to their post-traumatic stress symptom expression.^22,25–27 This suggests that the words describing individuals’ narrative of the traumatic event, representing the subjective, less-filtered experience of the trauma, could represent post-trauma adjustment, even before extensive psychological processing and meaning making has occurred.²⁷

Exciting developments in natural language processing (NLP) computational methods reveal that algorithms can analyze human language and extract insights similarly to how humans understand it, and in combination with machine learning (ML) models, they could be promising for informing the classification of psychiatric conditions.^28–30 Recent transformer-based NLP methods³¹ enable algorithms to achieve state-of-the art results in understanding contextual nuances of the language in written texts.³² NLP extracts and represents unstructured textual data as structured data that could then be used for generating ML classification models.³²

Despite the potential of analyzing the rich text in personal narratives for examining the experience of childbirth, there is a gap in knowledge on whether word usage in birth narratives could serve as markers of maternal mental health.^33,34 No study has examined the utility of childbirth narratives combined with advanced text-based computational methods and ML to inform the early identification of women endorsing traumatic stress in response to childbirth.

In this study, we collected short, unstructured narrative accounts of the personal and recent childbirth experiences from a total of 1,127 postpartum women. Using an NLP transformer-based algorithm and a developed ML classifier, we examined whether the text of narratives, alone, could be used to identify postpartum women with probable CB-PTSD.

Materials and Methods

Study design

This investigation is part of a research study concerning the childbirth experience and maternal psychological sequelae during the COVID-19 era.³⁵ Women who gave birth to a live baby in the last six months, and were 18+ years old were enrolled and provided information about their mental health and childbirth experience via an anonymous web survey. At the end of this survey, an option was provided to report one’s childbirth story. Narratives were collected at a mean timepoint of 2.50 ± 1.78 (range: 0.02–8.34) months following childbirth. Recruitment was during the period of 04/02/2020 to 12/29/2020, using hospital announcements, social media, and professional organizations. The project received exemption from the Partners Healthcare (Massachusetts General Brigham) Human Research Committee (PHRC).

The sample consists of 1,127 women who provided birth narratives, 1,111 of whom completed a PTSD symptom screen; of these, 995 (88.29%) provided written childbirth narratives of length 30+ words. On average, maternal age was 32 ± 4.43 years and gestational period was 38.98 ± 1.71 weeks. 53.1% of participants were primiparas, and 69.6% gave birth via vaginal delivery. Around 10% (n = 86) had a positive PTSD screen (PTSD Checklist for DSM-5, PCL-5 ≥ 31). Table 1 presents demographics and childbirth information for women with and without a positive screen. All fields in this table were collected via subject self-report.

Table 1.

Demographics and childbirth factors by childbirth-related posttraumatic stress disorder status

	CB-PTSD (n=86)		No CB-PTSD (n=909)		χ2	OR (95% CI)
	%	n	%	n
Maternal age
<35 years	67.4	58	68.2	620	0.02	0.97 (0.60–1.55)
Education
Less than bachelor’s degree	24.4	21	14.4	131	6.08*	1.92 (1.13–3.25)
Marital status
Not married or domestic partnership	10.5	9	5.9	54	2.71	1.85 (0.88–3.89)
Household income
<$100,000	49.4	42	39.3	354	3.31	1.51 (0.97–2.36)
Race and ethnicity
Black and/or Hispanic or Latino	15.3	13	7.4	67	6.45*	2.25 (1.19–4.27)
Prior mental health	52.3	45	32.2	293	14.14***	2.31 (1.48–3.60)
Primiparity	66.3	57	51.8	471	6.60**	1.83 (1.15–2.91)
Pain medication	82.6	71	76.5	695	1.61	1.45 (0.81–2.59)
Labor induction	57.0	49	45.8	416	3.93*	1.57 (1.00–2.45)
Obstetrical complications	54.7	47	24.1	219	37.46***	3.80 (2.42–5.96)
Sleep deprivation	84.2	64	65.5	539	11.04***	2.81 (1.49–5.29)
Mode of delivery
Vaginal	45.3	39	71.9	654	26.29***	0.32 (0.21–0.51)
Planned Cesarean	8.1	7	12.9	114	1.43	0.62 (0.28–1.37)
Unplanned Cesarean	46.5	40	15.5	141	50.74***	4.74 (2.99–7.50)
Premature delivery	14.0	12	5.7	52	8.85**	2.67 (1.37–5.23)
Sense of danger for self or newborn’s life	53.5	46	14.5	131	81.48***	6.80 (4.28–10.79)
Skin-to-skin	65.1	56	90.7	824	50.84***	0.19 (0.12–0.31)
Rooming in	72.1	62	92.8	831	41.40***	0.20 (0.12–0.34)
NICU admission	30.2	26	82.7	96	27.96***	3.65 (2.20–6.05)
Acute stress in childbirth	79.1	68	17.4	158	170.32***	17.96 (10.39–31.03)
	M	SD	M	SD
Mean postpartum study completion (months)	3.13	1.96	2.45	1.75

CB-PTSD: childbirth-related posttraumatic stress disorder with classification of CB-PTSD and No CB-PTSD based on the PTSD Checklist for DSM-5 cutoff of 31; Prior Mental Health: Responses to the question: “Before your most recent childbirth… Did you suffer from any mental health problems? Yes/No”. If Yes, provided checkboxes for the following conditions: Depression, Anxiety, PTSD, Postpartum Depression, and Other. If Other, please specify.”; Pain Medication: Responses to the question: “Did you receive any medication for pain? Yes/No”; Obstetrical complications: Binary response to the question: “Did you have any obstetric complications during labor or delivery? Yes/No”; Sleep deprivation : <6 hours of sleep the night before birth; Premature delivery <37 weeks gestation age; NICU : neonatal intensive care unit; Skin-to-Skin: Responses to the question: “Did you engage in skin-to-skin contact following the birth? Yes/No. If No, Was this due to COVID-19? Yes/No”; Rooming In: Responses to the question: “Was your baby in the room with you during your stay at the hospital? Yes/No. If No, Was this due to COVID-19? Yes/No.”; Sense of danger refers to during/immediately after childbirth, at least moderate degree; Acute stress refers to clinically significant immediate emotional and physiological responses to personal childbirth as measured by the Peritraumatic Distress Inventory (PDI ≥ 17).

OR: Odds ratio, 95% CI = 95% confidence interval.

^*p < .05,

^**p < .01,

^***p < .001.

Differences in sample sizes are due to missing data.

Measures

Narratives of childbirth

Narratives of childbirth were collected as written (typed) unstructured, open-ended accounts of each participant’s personal and recent childbirth experience. Narratives were collected in a free recall paradigm in which participants were instructed as follows: “Please provide a brief description of your recent childbirth experience in your own words, with a focus on the most distressing aspects of your experience, if applicable.” This focus on the most stressful aspects of the childbirth experience accords with the standard strategy for studying trauma sequelae in non-postpartum samples.^26,36

PTSD symptoms specific to childbirth

PTSD symptoms specific to childbirth were measured using the PTSD Checklist for DSM-5 (PCL-5),³⁷ the standard self-report measure to assess presence and severity of the 20 DSM-5 PTSD symptoms following an index traumatic event endorsed over the past month. Participants were instructed to report on their symptoms associated with their recent childbirth. The PCL-5 has strong correspondence with clinician diagnostic assessments and is used to determine provisional PTSD diagnosis (i.e., putative disease state, without formal diagnosis),³⁸ with a reported clinical cutoff of 31.³⁹ Reliability of the measure was high (Cronbach’s alpha = 0.91).

Machine learning modeling analysis

To analyze narrative text, we represented sentences (narratives) as dense, low-dimensional vectors (‘embeddings’), using the all-mpnet-base-v2 pre-trained Sentence-Transformers NLP model.⁴⁰ This model maps sentences and paragraphs to a fixed size of 768-dimensional dense vector. It fine-tuned Microsoft’s pretrained mpnet-base NLP model⁴¹ on a dataset of 1 billion sentence pairs to identify sentence similarity (essential to our developed method), and it was reported to produce the highest average performance⁴² on encoding sentences over 14 NLP tasks, compared with other NLP models. We used the Siamese network approach to identify semantically similar sentences, in terms of meaning; Appendix A provides details about this method.

We developed an ML model that utilizes the output (sentence embedding vectors) of the all-mpnet-base-v2 NLP model to identify CB-PTSD via narrative classification. The developed ML model was trained to classify childbirth narratives as markers of endorsement, or no endorsement, of CB-PTSD. Appendix A presents the 4 steps to build and test our model, and Appendix B provides sensitivity analysis of our model.

We labeled narratives associated with PCL-5 ≥ 31 as “CB-PTSD” (Class 1), and PCL-5 < 31 as “No CB-PTSD” (Class 0) (Step 1, Appendix A). We discarded short narratives (<30 words) to allow meaningful learning of word patterns,⁴³ resulting in the removal of 116 narratives (Step 2.1, Appendix A). Then, we balanced the dataset using down-sampling by randomly sampling the majority Class 0 to fit the size of the minority Class 1, resulting in 86 narratives in each class. We constructed the train and test datasets as described in Step 2.2 in Appendix A. We repeated this step (and the following steps) 10 times to capture different narratives for creating an accurate representation of Classes 0 and 1.

Next, we created three sets of sentence pairs using the train set. Set #1: All possible pairs of sentences (2,145) in Class 1 (CB-PTSD); Set #2: All possible pairs of sentences (2,145) in Class 0; and Set #3: Pairs of sentences (4,290), one randomly selected from Class 1 and another randomly selected from Class 0. We labeled Sets #1 and #2 as positive examples since they contained semantically similar pairs of sentences (either a pair of narratives of participants with or without CB-PTSD). We labeled Set #3 as negative examples since they contained pairs of non-semantically similar pairs of sentences. This data augmentation process produced 8,580 training examples generated from the train set (Step 3.1, Appendix A).

We mapped each narrative using the all-mpnet-base-v2 model into a 768-dimensional vector. We standardized these vectors by removing the mean, and scaling to unit variance. Finally, we computed the absolute element-wise difference between each of the 8,580 embedding vectors of pair of sentences u, v in Sets #1 to #3 of the train set (Step 3.1, Appendix A), such that z = (|emb(u) − emb(v)|) (Step 3.2, Appendix A).

Using the 8,580 calculated z vectors, we trained a deep feedforward neural network (DFNN) model to classify pairs of sentences in Sets #1 to #3 as semantically similar or not (Step 3.3, Appendix A). For training, we used the Keras Python library and constructed a DFNN with an input layer of 768 neurons, two hidden layers of 400 and 50 neurons, and an output neuron. All layers had a ReLU activation function, excepting the output neuron with a Sigmoid activation function. We used 150 epochs, applying the Adam optimizer with a learning rate of 3e⁻⁵, batch size of 64, and binary cross-entropy loss to monitor training performance. To avoid overfitting, we stopped training when there was no loss improvement for three consecutive epochs. We used 20% of the train dataset for validation during the training process.

Finally, we tested and compared the performance of our developed model against a baseline model using a 10-fold cross-validation (CV) (Step 4, Appendix A). As a baseline model, we fine-tuned the all-mpnet-base-v2 NLP model using the train dataset on a downstream task of classifying narratives into Class 0 or 1. We used the Sentence-Transformers Python library within HuggingFace Hub with the following parameters: learning rate=2e⁻⁵, batch size=16, epochs=50, weight decay=0.001. We evaluated our model performance using (i) the F1 score, which is a measure integrating Precision (positive predictive value) and Recall (sensitivity); and (ii) Area Under the Curve (AUC). We also created a calibration curve, and a Receiver Operating Characteristic (ROC) curve (Appendix C).

Results

Descriptive

Following the data processing (Steps 1 and 2, Appendix A), for Class 1 (CB-PTSD) and Class 0 (No CB-PTSD), mean and median word count (WC) were 191.91 and 142, and 154.61 and 106, respectively. A t-test analysis revealed that participants of the CB-PTSD class used more words to describe their childbirth experiences than those of the No CB-PTSD class (t = 2.30, df = 111.99, p = 0.02) (Figure 1).

Figure 1.

Number of words in childbirth narratives by childbirth-related PTSD status. Boxplots display word count (WC) in narratives for CB-PTSD (Class 1, pink) and No CB-PTSD (Class 0, light blue) based on PTSD Checklist for DSM-5 (PCL-5) cutoff of 31. Dots are data points (narratives’ WC) shifted by a random value. The mean WC for Class 1 and Class 0 was 191.91 and 142, and the median WC was 154.61 and 106, respectively. Participants of the Class 1 used more words to depict their narrative than those of Class 0 (t = 2.30, df = 111.99, p = 0.02).

Machine learning modeling analysis

The results of applying the baseline model to the test set and our model are presented in Table 2. Our developed model outperformed the baseline model on the tested metrics of Area Under the Curve (AUC), F1 score, Sensitivity, and Specificity.

Table 2.

Comparison of the average (10 different seeds) performance classification results of the developed model vs. a baseline model. Both models use exclusively text features to identify childbirth-related PTSD.

Model	AUC*	F1 Score**	Sensitivity	Specificity
Baseline model	0.53	0.52	0.66	0.41
The developed model	0.75	0.76	0.80	0.70

^*A model with an AUC of 0 suggests no ability to diagnose patients, and 1 a perfect diagnosis.

^**F1 score ranges between 0 to 1 (a perfect classification).

The baseline model’s results emphasize the problem of training an ML classifier with a small number of examples. In contrast, our model was able to overcome this problem by using 8,580 training examples, outperforming the baseline model. As reported in Table 2, and in particular, regarding the F1 score (0.76) and AUC (0.75), our model for CB-PTSD classification derived from birth narratives achieved overall good performance.

Word category analysis

To examine usage of specific word categories in birth narratives and their potential relation to CB-PTSD status, we used the Linguistic Inquiry and Word Count (LIWC) software,⁴⁴ which uses different validated dimensions to classify words into categories. It compares a word from the natural text input to a dictionary of pre-defined words, and classifies the identified word into a predefined dimension.⁴⁴

We examined the frequency of word categories in childbirth narratives that were previously shown to represent trauma narratives of individuals with PTSD.^26,45 Using LIWC, we examined the frequency of: ‘Affect’, ‘Anger’, ‘Anx’, ‘Bio’, ‘Body’, ‘Cause’, ‘Cogproc’ (cognitive processes), ‘Death’, ‘Feel’, ‘Filler’, ‘Health’, ‘Hear’, ‘I’, ‘Insight’, ‘Negemo’ (negative emotions), ‘Percept’, ‘Posemo’ (positive emotions), ‘Pronoun’, ‘Sad’, ‘See’, ‘We’, and ‘You’. A Wilcoxon test for differences in word frequency revealed that participants of the CB-PTSD class used fewer positive emotions, and more negative emotions as well as body- and death-related words, compared with the No CB-PTSD class (Figure 2).

Figure 2.

Frequency of words in childbirth narratives by childbirth-related PTSD status. Distribution of word frequencies (LIWC value) for CB-PTSD (Class 1, pink) and No CB-PTSD (Class 0, light blue) based on PTSD Checklist for DSM-5 cutoff of 31. Significant results of Wilcoxon rank sum tests with continuity correction for differences in word frequency between classes was found for ‘anger’: W=35,005.50, p=1.32e-02; ‘bio’: W=28,710.50, p=4.63e-05: ‘body’: W=27,438.50, p=2.00e-06; ‘death’: W=34,538, p=3.48e-05; ‘health’: W=33,443, p=2.66e-02; the pronoun ‘I’: W=33,891, p=4.14e-02; ‘posemo’ (positive emotions): W=49,581.50, p=3.71e-05; ‘sad’: W=31,017, p=8.90e-04; and ‘see’: W=32,402, p=2.30e-03. X-axis label “I” is the first-person pronoun.

Comment

Principal findings

This study shows that a neural network machine learning (ML) model trained on natural language processing (NLP) features of free text (narratives) has the potential to identify women with CB-PTSD following childbirth, based on short, unstructured personal childbirth written narrative accounts. This simple data collection method appears feasible and efficient for collecting information from postpartum women during a sensitive time period, and may overcome inherent barriers of relying on medical record data to identify at-risk women.^46–48 In our model, 80% (sensitivity) of women who meet CB-PTSD criteria could be accurately identified based on word usage in their narratives, and 70% (specificity) of those not endorsing the condition could be identified as such.

Results in the context of what is known

To our knowledge, this is the first study to use childbirth narratives accounts and state-of-the-art NLP algorithms combined with ML models for the identification via classification of a maternal mental health condition in general.⁴⁹ Research using ML models for the classification of CB-PTSD is largely lacking. Only a few studies have tested the utility of ML models for CB-PTSD identification.^50,51 While our model’s performance is comparable to the reported models, those models are informed by relatively extensive data such as information derived from medical records⁵⁰ and/or structured questionnaires.^50,51 In contrast, the use of freely generated childbirth narratives may not only have the advantage of being a more accessible data collection method, but also entails self-disclosure and narrative construction, which both have positive implications in the processing of traumatic events and facilitating psychological adaptation.^52–55

Clinical implications

Although early and mass screening for CB-PTSD would likely improve diagnosis rates and facilitation of treatment, there are no recommended medical protocols for CB-PTSD screening in hospitals and health clinics. The opportunity to screen women when they are still in contact with obstetrical providers is important, as such contact appears much more difficult to establish later, when disorders can become chronic and often comorbid,⁶ and hence, more difficult to treat.⁵⁶ Establishing the potential accuracy of non-invasive and low-cost data collection based on childbirth narratives for the identification of women with CB-PTSD could serve as an important first step to complement more extensive clinical assessments and biologically-oriented methods.⁵⁷ Collection of short written childbirth accounts could be done remotely, potentially ahead of a medical visit, so that care could be provided to those identified as being at high risk for CB-PTSD. Our strategy of assessing childbirth narratives permits the possibility of introducing an assessment with minimal burden during an acute period of rapid physiological and psychological adjustment. NLP analysis of subject-generated text is starting to be used for disease diagnosis in research contexts,^58,59 and its application to detect CB-PTSD is novel. Our strategy may have the potential to be readily implemented in routine obstetrical care, and may warrant developing a commercial model to facilitate widespread adoption. Additionally, the racial and ethnic disparities related to childbirth-associated trauma¹² remain to be explored in detail. Our model has the potential to improve risk assessment of CB-PTSD, particularly in minority populations, by accounting for these considerations.

Research implications

We applied an advanced NLP model for analyzing the text in personal childbirth narratives. Although NLP has been increasingly used to analyze free text in electronic medical records (EMR) documented by medical staff, the application of NLP to birth narratives written by the patients themselves for identification of CB-PTSD has not previously been done. We report an AUC of 0.75 that accords with the performance of other psychiatric studies of non-postpartum classifications using NLP models.⁶⁰ The uniqueness of the method applied here is the development of a pairwise sentence similarity classification model that uses a pre-trained Sentence-Transformers NLP model to embed narratives, and uses these embeddings to train a neural network classifier for CB-PTSD identification. Our results suggest that computational models that can accurately understand the context of the language, similarly to how humans do, are promising for detecting maternal psychiatric morbidity from narrative birth-related text. Future studies are needed to replicate our results and can potentially examine the utility of briefer passages of childbirth narratives. The accuracy of our model could be further improved by (1) training an NLP model (instead of using a pre-trained model) to learn the unique language used in CB-PTSD narratives and better represent narratives using embedding vectors, and (2) including additional non-textual data types in the model, such as EMR data.

Strengths and limitations

This study presents the first use of NLP to identify a maternal mental health condition from women’s personal written accounts of their childbirth experience collected remotely. Our model demonstrated good performance; for future work, combining such narratives with information in patients’ medical records, and physiological and emotional responses to childbirth, may enhance the accuracy in forecasting the maternal mental health outcome following traumatic experiences of childbirth.

Several limitations to this work are worth noting. Electronic medical record data were not used; future work could integrate this data source with the patient self-reports employed in our study. Although we used the well-validated PTSD checklist for DSM-5 that corresponds strongly with diagnostic measures⁶¹ to determine CB-PTSD, some participants were assessed before CB-PTSD can be confirmed, and clinician assessments were not performed. Future work should consider secondary outcomes, e.g., treatment need, to establish more robust findings. The samples largely represent middle-class American women, warranting replication of the work in more diverse populations. The model was developed and tested on the collected dataset; however, we used a pre-trained NLP model to represent narratives as vectors, which may not fully capture the language nuances in childbirth narratives; this necessitates future work in training a new NLP model on a larger narrative dataset.

Conclusions

In summary, this study is a proof of concept of the potential utility of the text in childbirth narratives, alone, to inform the detection of early signs of maternal PTSD following childbirth using state-of-the-art NLP and ML models. As psychiatric morbidity in the transition to motherhood remains a public health concern,^62,63 more research is needed to guide the development of tools for the accurate and early screening of women likely to endorse a mental illness following childbirth.

AJOG at a Glance:

A. Why was this study conducted?

Maternal PTSD following traumatic childbirth is an undertreated and underdiagnosed postpartum psychopathology.

Research to inform the accurate identification of women at high risk of childbirth-related PTSD is lacking.

B. What are the key findings?

Women with childbirth-related PTSD construct longer accounts of their childbirth experience and use more negative emotions and death-related words in their stories.

Natural language processing and machine learning algorithms applied to unstructured, short childbirth narratives produce a computational model that can identify women with childbirth-related PTSD with relatively good accuracy.

C. What does this study add to what is already known?

The use of words in childbirth narratives alone may inform a low-cost, low-burden screening tool for early identification of women at risk for childbirth-related PTSD.

Applying a data-driven approach combining machine learning and advanced natural language processing methods for free text analysis may be useful in predicting maternal psychiatric health outcomes.