Egocentric network maps—a visual depiction of the personal connections in one’s life—are a valuable tool for communicating, understanding, and analyzing the social resources that foster positive medical and other outcomes. Gathering this data is an in-depth process, often conducted through surveys or interviews. Given concerns of participant burden and cognitive load, interviews to gather this information are often considered the gold standard. Yet, such methods generate immense amounts of data that must been cleaned and mined for network mapping.
As part of the AQUASS Research Lab and SNRC in the Department of Emergency Medicine, Katelyn Carlsen—an undergraduate in the Judy Genshaft Honors College who plans to pursue an MD/PhD and advance research and practice in neurology, is working alongside a team to address the time-intensive nature of manual data extraction from interview transcripts by applying Natural Language Processing techniques, a subset of Artificial Intelligence (AI). The idea is to allow for the extraction and analysis of qualitative and quantitative data much faster than traditional techniques to collect personal network information, without sacrificing accuracy.
Social network science has identified multiple social network attributes that have been linked to social and medical outcomes. Therefore, participants are typically asked to list their social network members (alters), describe relationships among those individuals or whether they know each other (network density), in what ways/roles they know that person (role multiplexity), for how long they have known them (strength of ties), the types of resources they provide, their attributes (e.g., demographics), and so on. Thus, analysis of this data can be cumbersome due to the sheer volume and multidimensional nature of it.
The effort utilizes data from the USF hyperbaric oxygen therapy study, a randomized controlled trial (RCT). The trial explores the efficacy of hyperbaric oxygen as a treatment for Veterans who have sustained mild to moderate traumatic brain injury/injuries (TBI) by delivering a high concentration of oxygen in a pressurized environment. Social network mapping is part of the medical anthropology HBOT sub-study, which investigates the information and attitudes that participants hold throughout their treatment, noting that such influences are often informed by participants’ social relationships. As these factors may impact experience throughout the study, it is important to carefully examine them through the analysis of qualitative and quantitative social networks along with data collected through semi-structured interviews.
To illustrate how alters in a participant’s social circle may affect information, attitudes and even ability to participate in the trial, the team is constructing egocentric network diagrams. These place participants as focal (ego) and illustrate their immediate social connections. Analysis of these networks can provide insight into how information, attitudes, and resources move through individuals within a network.
Carlsen has two years of experience analyzing data using machine learning techniques. At this year’s USF Health Research Day in March, Carlsen presented the initial efforts, illustrating how accurately the current NLP model can identify features of egocentric networks from qualitative data. Her multi-step analytical pipeline begins with transcribing audio recordings, followed by tokenization (breaking text into smaller units) and parts of speech tagging. After this, processed text moves through a rule-based classification system that utilizes Named Entity Recognition (NER) and Sentiment Analysis.
Rule-based classification operates by matching text against a predefined set of linguistic rules and assigns categories accordingly. In this case, categories include alter attribute information such as the length of relationship, the type of support provided, and other demographic information. NER allows text to be categorized as entities (such as names, ordinal/cardinal numbers, organization, etc), based on a pre-trained machine learning model. This allows for the capture of relevant alter name identification. Sentiment analysis allows for text to be categorized as positive, neutral, or negative based on the emotional tone of words, allowing for an automated surface-level qualitative analysis.
Preliminary findings demonstrate that NLP holds considerable promise for social network data extraction. The methods were able to correctly identify attributes such as gender and age for a majority of alters, and showed potential in identifying other characteristics, such as types of support.
Looking ahead, Ms. Carlsen highlights several opportunities to improve this approach. Incorporating machine learning techniques and refining interview protocols, particularly in longitudinal research, could enhance qualitative accuracy in both manual and automated analyses. Future work may also include adding coreference resolution to better track alters through interview data, allowing the machine to track words and text that are referring to specific individuals, thereby enhancing the accuracy of data extraction. Finally, developing more anthropologically informed sentiment analysis through the custom training of annotated examples may help to provide data more meaningfully attuned to the medical anthropology study goals.
Showing the relevance of this work across disciplines, Ms. Carlsen also presented her work at USF’s 2026 AI+X Symposium, an event showcasing research that demonstrates interdisciplinary AI applications. While USF Health Research Day provided an audience with a broad background in health, this venue provided an opportunity to focus on the technical and methodological details of the work and receive feedback from experts in the field of AI. The revised work was also presented at the International Network for Social Network Analysis annual Sunbelt meeting in Daytona Beach in June 2026 (poster below).