Presentation
PART 2: Bridging conversation islands to connect healthcare: Introducing unique co-occurring word networks to find distinct themes
Event Type
Oral Presentations
TimeThursday, June 9th3:00pm - 3:30pm EDT
Location
DescriptionIntroduction
Safety, human factors, and systems engineering draw from many disciplines. So many in fact it is a challenge, if even possible, to translate between the specialty areas that have emerged. Meister (1995) suggests the discipline of human factors has become fractionated because of specialization. In an age of hyperspecialization, it is probably fair to say that about many disciplines. Disciplines, efforts, and educational programs have emerged with the specific focus to address interdisciplinary collaboration and sharing. Described in this paper is another approach to sharing, one for peering into disciplines and specialties in search of distinct work and technologies. Identifying emerging work systems and technologies, earlier on if even at all, is critical to the prospective design of human-centered and safer healthcare systems.
If the discipline of Human Factors seeks to understand how people interact with their surroundings, then language is key to how we interact and how we explain that interaction. If Human Factors seeks to understand how people interact from physical, cognitive, and cultural perspectives, then it is precisely as important to understand what is missing from a community’s perspective of their interactions and how they are not interacting with work technologies and environments.
This paper describes a new text analytics approach for finding information about technologies, ways of doing work, and risks that have the attention of one community and seemingly not another. Imagine this isolated information as a conversation island, something a community is interested enough in writing down, but it is not shared with or discussed by another community.
To build a bridge between islands, texts from two or more communities are examined drawing out distinct themes that each may find useful. To find distinct topics, a simple approach is described for filtering unique words from these bodies of text and assembling these words into network visualizations for exploration. This approach potentiates the discovery of distinct themes and subsequent validation efforts. Once substantiated, if necessary, findings can be used to facilitate awareness and inform various design, usability, safety, and risk mitigation efforts.
Approach and Examples
To assemble unique co-occurring word (u-coord) networks is theoretically a simple task. The first step requires identifying and collecting texts associated with the domain of interest that is easily separated into two groups for extraction, cleansing, processing, and analysis. These groups must share something in common, but also be differentiable. The strategy outlined in this paper has been applied for multiple purposes (table 1).
Table 1: List of applications
• Design technical requirements
• Identify emerging technologies
• Discover opportunities in Human Factors
• Inform human-centered design
• Facilitate language awareness
• Review patient safety reports
• Compare usability findings
• Formulate instructional design
Consider, for example, designing requirements to guide the purchase of some category of medical device. For formulating these requirements, multiple information sources are reviewed such as applicable standards, instructions for use, product specification documents, and safety reports, to list a few. This requires wading through, making sense of available information, and convening a team of subject matter experts that meet periodically to make decisions and refine requirements. As a supplement to human review, using computational means to compare the text from the applicable standards with text from safety reports can illuminate important unidentified usability features. In this example, one latent theme identified was a cluster of usability issues relative to how some devices displayed information about weight. Information and visualizations can be shared with the team interactively, inform decisions going forward, and facilitate the purchase of safer designs.
Once texts have been identified, collected, and separated in two groups along a differential, the bodies of text are tokenized, token frequency calculated, and the two lists merged with an outer join. During tokenization, unique words, phrases, or ngrams are extracted and are often referred to as tokens (Manning & Schutze, 1999). In this paper, words and tokens will be used synonymously. Tokens unique to each list are extracted into two lists: one list of unique tokens from each group. These preprocessing steps are used to create and align the two lists to probe for unique word co-occurrences. A manual review helps identify unique terms and surrounding context as a first step in conceptualizing distinctiveness. However, individual lists can be cumbersome and a challenge for manual review.
To create more usable representations of the distinct terms in each group a simple trick is used for visualizing latent themes. To begin this task, the unique co-occurrences are assembled into a list of co-occurrences with their frequencies. Then the magic begins, an interactive network is created from each list of unique co-occurrences (Figure 1, image not included in proposal). This step connects unique words found together in the narrative text to provide network clusters helping in the interpretation of word use within the context of others in proximity.
Unlike the visual seen through the eyes of a concordance viewer, words found in the contrasting text are missing from the network representation. The reviewer can interact with these networks and explore in greater context by connecting with and reviewing words within their narrative text. Comparable to a traditional hermeneutic approach to interpretation by moving between the parts and the whole with a contemporary twist by adding computational resources and visualization techniques to design a crescendo of information (Figure 2).
Unique Word – Unique Word Network – Narrative Text
Less---------------------Information-------------------More
Less-------------------------Work-----------------------More
Figure 2: Level of information and work as a function of text volume
After selecting distinct themes of interest, the last step is to substantiate the findings, if necessary. This step may be skipped if the purpose or available resources make it unnecessary or unfeasible. To perform this step, the reviewer uses an online academic or library search engine to enter different combinations of the words from each individual word network. Adding the discipline or domain of interest to this query may also be beneficial.
If this search provides results, then documents can be retrieved, electronically searched for the words of interest, and context reviewed. If it is confirmed by review that these newly retrieved documents are like the theme in question, then the theme in question may not be as distinct as initially thought. To speed up the review process, a concordance or corpus viewer can be used to scan newly retrieved documents. This step does not provide complete confidence however, it can offer a level of confidence, as to whether or not gaps exist in the domains of interest.
Figure 3 schematically outlines the process for designing, exploring, and substantiating u-coord networks. Several examples of previous work and the benefits of this approach will be provided in the presentation and conference proceeding’s paper (Arnold & Fuller, 2017; Arnold & Fuller, 2018; Arnold, unpublished).
• Identify purpose and select representative collection of texts
• Separate texts into two or more groups
• Note domain or disciplines and similarities and differences
• Extract and clean texts
• Process and tokenize texts
• Create word frequency tables
• Join tables and sort
• Identify unique words
• Create unique word co-occurrence frequency tables
• Create and review u-coord networks
• Select u-coord networks of interest
• Investigate and substantiate with online query if necessary
Figure 3: A process diagram for designing and substantiating u-coord networks will be provided
Safety, human factors, and systems engineering draw from many disciplines. So many in fact it is a challenge, if even possible, to translate between the specialty areas that have emerged. Meister (1995) suggests the discipline of human factors has become fractionated because of specialization. In an age of hyperspecialization, it is probably fair to say that about many disciplines. Disciplines, efforts, and educational programs have emerged with the specific focus to address interdisciplinary collaboration and sharing. Described in this paper is another approach to sharing, one for peering into disciplines and specialties in search of distinct work and technologies. Identifying emerging work systems and technologies, earlier on if even at all, is critical to the prospective design of human-centered and safer healthcare systems.
If the discipline of Human Factors seeks to understand how people interact with their surroundings, then language is key to how we interact and how we explain that interaction. If Human Factors seeks to understand how people interact from physical, cognitive, and cultural perspectives, then it is precisely as important to understand what is missing from a community’s perspective of their interactions and how they are not interacting with work technologies and environments.
This paper describes a new text analytics approach for finding information about technologies, ways of doing work, and risks that have the attention of one community and seemingly not another. Imagine this isolated information as a conversation island, something a community is interested enough in writing down, but it is not shared with or discussed by another community.
To build a bridge between islands, texts from two or more communities are examined drawing out distinct themes that each may find useful. To find distinct topics, a simple approach is described for filtering unique words from these bodies of text and assembling these words into network visualizations for exploration. This approach potentiates the discovery of distinct themes and subsequent validation efforts. Once substantiated, if necessary, findings can be used to facilitate awareness and inform various design, usability, safety, and risk mitigation efforts.
Approach and Examples
To assemble unique co-occurring word (u-coord) networks is theoretically a simple task. The first step requires identifying and collecting texts associated with the domain of interest that is easily separated into two groups for extraction, cleansing, processing, and analysis. These groups must share something in common, but also be differentiable. The strategy outlined in this paper has been applied for multiple purposes (table 1).
Table 1: List of applications
• Design technical requirements
• Identify emerging technologies
• Discover opportunities in Human Factors
• Inform human-centered design
• Facilitate language awareness
• Review patient safety reports
• Compare usability findings
• Formulate instructional design
Consider, for example, designing requirements to guide the purchase of some category of medical device. For formulating these requirements, multiple information sources are reviewed such as applicable standards, instructions for use, product specification documents, and safety reports, to list a few. This requires wading through, making sense of available information, and convening a team of subject matter experts that meet periodically to make decisions and refine requirements. As a supplement to human review, using computational means to compare the text from the applicable standards with text from safety reports can illuminate important unidentified usability features. In this example, one latent theme identified was a cluster of usability issues relative to how some devices displayed information about weight. Information and visualizations can be shared with the team interactively, inform decisions going forward, and facilitate the purchase of safer designs.
Once texts have been identified, collected, and separated in two groups along a differential, the bodies of text are tokenized, token frequency calculated, and the two lists merged with an outer join. During tokenization, unique words, phrases, or ngrams are extracted and are often referred to as tokens (Manning & Schutze, 1999). In this paper, words and tokens will be used synonymously. Tokens unique to each list are extracted into two lists: one list of unique tokens from each group. These preprocessing steps are used to create and align the two lists to probe for unique word co-occurrences. A manual review helps identify unique terms and surrounding context as a first step in conceptualizing distinctiveness. However, individual lists can be cumbersome and a challenge for manual review.
To create more usable representations of the distinct terms in each group a simple trick is used for visualizing latent themes. To begin this task, the unique co-occurrences are assembled into a list of co-occurrences with their frequencies. Then the magic begins, an interactive network is created from each list of unique co-occurrences (Figure 1, image not included in proposal). This step connects unique words found together in the narrative text to provide network clusters helping in the interpretation of word use within the context of others in proximity.
Unlike the visual seen through the eyes of a concordance viewer, words found in the contrasting text are missing from the network representation. The reviewer can interact with these networks and explore in greater context by connecting with and reviewing words within their narrative text. Comparable to a traditional hermeneutic approach to interpretation by moving between the parts and the whole with a contemporary twist by adding computational resources and visualization techniques to design a crescendo of information (Figure 2).
Unique Word – Unique Word Network – Narrative Text
Less---------------------Information-------------------More
Less-------------------------Work-----------------------More
Figure 2: Level of information and work as a function of text volume
After selecting distinct themes of interest, the last step is to substantiate the findings, if necessary. This step may be skipped if the purpose or available resources make it unnecessary or unfeasible. To perform this step, the reviewer uses an online academic or library search engine to enter different combinations of the words from each individual word network. Adding the discipline or domain of interest to this query may also be beneficial.
If this search provides results, then documents can be retrieved, electronically searched for the words of interest, and context reviewed. If it is confirmed by review that these newly retrieved documents are like the theme in question, then the theme in question may not be as distinct as initially thought. To speed up the review process, a concordance or corpus viewer can be used to scan newly retrieved documents. This step does not provide complete confidence however, it can offer a level of confidence, as to whether or not gaps exist in the domains of interest.
Figure 3 schematically outlines the process for designing, exploring, and substantiating u-coord networks. Several examples of previous work and the benefits of this approach will be provided in the presentation and conference proceeding’s paper (Arnold & Fuller, 2017; Arnold & Fuller, 2018; Arnold, unpublished).
• Identify purpose and select representative collection of texts
• Separate texts into two or more groups
• Note domain or disciplines and similarities and differences
• Extract and clean texts
• Process and tokenize texts
• Create word frequency tables
• Join tables and sort
• Identify unique words
• Create unique word co-occurrence frequency tables
• Create and review u-coord networks
• Select u-coord networks of interest
• Investigate and substantiate with online query if necessary
Figure 3: A process diagram for designing and substantiating u-coord networks will be provided