NATURAL LANGUAGE PROCESSING AS A TOOL FOR SCREENING DEPRESSION IN POSTPARTUM WOMEN SUFFERING FROM INTIMATE PARTNER VIOLENCE
BACKGROUND: Intimate partner violence (IPV) and depression are well-documented psychosocial stressors faced by perinatal women, with estimates of 3.9 to 8.3% who experience intimate partner violence and 7.1 to 12.7% who experience depression. Each presents immediate as well as long-term health risks to both the woman and her newborn. The impact of IPV and depression is often cumulative as each can exacerbate the consequences of the other. Psychosocial screening within clinical settings is common but faces multiple barriers to implementation. Alternative methods of detection, such as the use of natural language processing, which involves the analysis and synthesis of linguistic data, would be a valuable clinical tool.
OBJECTIVE: The goal of the current study is to examine the effectiveness of natural language processing as a tool for screening depression in postpartum women suffering from intimate partner violence.
METHODS: Three-hundred-twenty-six postpartum women were screened for IPV via a phone interview using three questions for current or past emotional or physical abuse. Sixty-four women screened positive, and 31 of these interviews were subsequently completed, taped, and transcribed. The women were screened three times for depression using the Edinburgh Postnatal Depression Scale (EDPS). The National Research Council Word-Emotion Association Lexicon was used to establish percentage of sentiment-associated words within each transcribed interview. Resampling-based permutation tests were performed to assess differences in proportions of sentiment words between those with and without depression. Principal component analysis was used to reduce dimensionality combining the significant sentiment categories into one principal component. Logistic regression was used to associate the principal component with depression. ROC analysis was performed to assess the diagnostic accuracy compared to EDPS.
RESULTS: Permutation hypothesis tests revealed that women with major depressive disorder used a greater proportion of words associated with negativity (p=.0004), anger (p=.0057), disgust (p=.0007), fear (p=.0009), and sadness (p<.0001). These five sentiment categories loaded onto the first principle component explaining 73.10% of the variance. This combined score was significantly associated (p=.0078) with depression. The proposed natural language processing tool is 100% sensitive in its ability to detect depression screened by EDPS. Adjusting for the imperfect standard of EDPS, the perceived sensitivity to detect depression is 92.7%, and the perceived specificity is 24.1%.
CONCLUSION: Natural language processing has the potential to develop into an alternative method of detecting sensitive psychosocial problems. Natural language processing can be used as a reliable tool to detect depression in postpartum women suffering from intimate partner violence, although, further natural language processing methods should be explored to decrease the number of false negatives.