Introduction. We systematically analyzed the most commonly used narrative note formats and content found in primary and specialty care visit notes to inform future research and electronic health record (EHR) development. Methods. We extracted data from the history of present illness (HPI) and impression and plan (IP) sections of 80 primary and specialty care visit notes. Two authors iteratively classified the format of the sections and compared the size of each section and the overall note size between primary and specialty care notes. We then annotated the content of these sections to develop a taxonomy of types of data communicated in the narrative note sections. Results. Both HPI and IP were significantly longer in primary care when compared to specialty care notes (HPI: n = 187 words, SD[130] vs. n = 119 words, SD [53]; p = 0.004 / IP: n = 270 words, SD [145] vs. n = 170 words, SD [101]; p < 0.001). Although we did not find a significant difference in the overall note size between the two groups, the proportion of HPI and IP content in relation to the total note size was significantly higher in primary care notes (40%, SD [13] vs. 28%, SD [11]; p < 0.001). We identified five combinations of format of HPI + IP sections respectively: (A) story + list with categories; (B) story + story; (C) list without categories + list with categories; (D) list with categories + list with categories; and (E) list with categories + story. HPI and IP content was significantly smaller in combination C compared to combination A (-172 words, [95% Conf. -326, -17.89]; p = 0.02). We identified seven taxa representing 45 different types of data: finding/condition documented (n = 14), intervention documented (n = 9), general descriptions and definitions (n = 7), temporal information (n = 6), reasons and justifications (n = 4), participants and settings (n = 4), and clinical documentation (n = 1).
Conclusion. We identified commonly used narrative note section formats and developed a taxonomy of narrative note content to help researchers to tailor their efforts and design more efficient clinical documentation systems.
Primarily driven by financial incentives provided by the Meaning Use program, the U.S. health system has reached unprecedented electronic health record (EHR) adoption 1 , which was achieved through large-scale adoption of commercial EHRs 2 . These systems were adopted before facing a thorough redesign 3 and before accumulation of compelling evidence of their full impact 4 . As a result, new unintended consequences have emerged at different levels of the U.S. health system 5 . Some notorious limitations of current systems include confusing interfaces 6 , excessive, overzealous alerts and reminders 7 , and bloated clinical notes 8 . The latter is particularly challenging because the clinical documentation that is essential in facilitating memory and recall and to enable understanding and care coordination 9 , has become an increasing source of frustration among clinicians 10 . Clinical notes play an important role in the so called EHR-associated documentation burden as they contribute to information overload 11 . In addition, notes created in the U.S. digital health system are significantly longer than similar documentation in other developed countries 12 , are not created to increase clinicians’ situational awareness 13 , often contain redundant information 14 , and in some cases may never be read despite containing information relevant to patient care 15 . To compensate for the need to enter narrative data manually, clinicians frequently create their notes by using the patient’s previous note, a remediable practice known as “copy-and-paste” 16 , which facilitates immediate data entry but further aggravates the documentation burden, and compromises note quality and patient safety 17 .
To address EHR limitations and their impact on clinical documentation, informatics leaders and researchers have proposed multiple solutions such as improving EHR design, interoperability, smarter clinical decision support (CDS) systems, transferring some data entry to patients 18 , and more recently, adoption of speech recognition and other computational methods 3 . We believe that these solutions, if implemented in isolation, may not be sufficient because they will be applied without a formal representation of relevant data about patients’ care context. We define patients’ care context as contextual information about the patient’s health and care (e.g., patient reports that his symptoms are getting worse), the underlying reasons behind a clinician’s decisions (e.g., a new imaging test is needed because the patient’s symptoms are worse), or a clinician’s interpretations (e.g., I believe the patient is not eligible for advanced therapy). By not being able to access relevant information considered by clinicians in their care decisions, irrelevant alerts continue to be fired 7 , data-entry consumes precious time that would otherwise be dedicated to direct patient care 19 , and EHR navigation forces clinicians to access isolated EHR components, which includes the tedious tasks of finding and reading long notes to search for data that are not always documented or cannot be found 20 .
We hypothesize that improving the EHR and particularly clinical documentation systems, will demand efforts in two important research areas: (1) creation of a formalism to represent patients care context data in computable form (modeled, structured, and coded); (2) a better understanding of the narrative content communicated in clinical notes. In previous studies we have developed a formalism to represent patients’ care context as a collection of subject-predicate-object or concept-relationship-concept tuples (e.g., patient-has_intervention-amoxicillin or amoxicillin-has_reason-allergic to ciprofloxacin). These tuples were formulated through a series of studies reporting clinical and contextual concepts and semantic relationships elicited from the literature 21-22 and in annotations of clinical notes and patient-physician spoken communications during clinical visits 23 . In the latter, we have suggested several applications of the tuples to improve EHR navigation, CDS, learning health systems and data entry at the point of care.
Although several studies have assessed clinicians’ perceptions of clinical documentation systems from multiples perspectives such as the purposes of clinical notes, documentation of clinicians’ reasoning in their notes, preferred methods for note creation and preferred methods for note reading/retrieval 24 , current literature does not provide a systematic assessment of the content communicated in narrative note sections. Therefore, to augment our findings from previous studies 21-23 and inform future research and EHR development, in the present study we explore the types of narrative data communicated in clinical notes and the most common formats used to create narrative note sections.
Settings and Study Design. We conducted a descriptive study to identify the most commonly used note section formats and types of data communicated in narrative note sections. We extracted data from the note sections history of present illness (HPI) and impression and plan (IP) from a random sample of 80 primary and specialty care visit notes written between 03/01/2019 and 06/30/2019 at the University of Alabama-Birmingham (UAB) primary and specialty care clinics. Forty primary care notes were selected from the specialty internal medicine and 40 specialty care notes were selected from the specialties ear, nose, and throat (ENT) (20 notes) and urology (20 notes). We choose the note sections HPI and IP because they are more likely to contain the expressivity of patients’ care context, which is not typically found in more structured note sections such as “medication list” or “laboratory results.” We chose the specialties internal medicine, ENT and urology because they provide a wide range of concepts ranging from common infections and chronic diseases to complex surgeries, thereby increasing their generalizability. The UAB Institutional Review Board approved the study.
Identification of note section format and size. One of the authors (TKC) extracted the content from HPI and IP sections of the selected notes and replaced any identifiable information with the acronym PHI (protected health information). Two authors (TKC and PID) iteratively classified the content of the sections of each note as story (i.e., two or more coherent sentences that form a paragraph - Figure 1 -A), list with categories (i.e., list of sentences that do not form a coherent paragraph, but are organized by category - Figure 1-B) and list without categories (i.e., list of sentences that neither form a coherent paragraph nor are organized by category - Figure 1-C). When two formats were found in the same note section (e.g., story and list with categories), we classified the section using the predominant format (i.e., the larger one by number of words). We calculated the total number of words for each note (the entire note), and the total number of words in each narrative note section (HPI and IP). Next, we identified the most common combinations of note formats used. We then performed the following comparisons: (1) the size of each note section (HPI and IP) between primary and specialty care notes; (2) the combined size of HPI and IP between primary and specialty care notes; (3) the total note size between primary and specialty care notes; (4) the proportion of content in HPI and IP in relation to the total note size between primary care and specialty care notes; and (5) the size of the different combinations of format for HPI and IP across the whole sample.
Classification of note content. We were unable to find a publicly available classification of narrative note content, and therefore, we developed a coding scheme with types of data that are likely to be communicated in narrative note sections (e.g., history of a finding or intervention) based on concepts represented in our collection of note content tuples previously reported 23 . This initial scheme was used by two authors (TKC and PID) who independently annotated the de-identified content of HPI and IP sections of a random sample of five notes, and then met in person to modify the initial coding scheme by adding, eliminating, or merging codes via consensus. These authors then used the updated coding scheme to annotate a random sample of 15 notes (five from each specialty - internal medicine, ENT, and Urology). We calculated inter-rater agreement and upon satisfactory agreement (> 80% agreement and K >.80) we split the remaining sample between the two authors. If no satisfactory agreement was achieved, another sample of 15 notes was annotated. This process was repeated until a satisfactory agreement was achieved or the whole sample was annotated. The coding scheme was modified as needed via consensus after each round. Annotations were performed at the sentence level with concepts (e.g., findings, procedures, orders) annotated individually, except when similar concepts were mentioned as a sequence in a sentence. For example, in the sentence “Patient has history of orbital cellulitis and bilateral nasal congestion,” the entire sentence was annotated as “history of finding/condition reported.” After annotating the 80 notes in our sample, the coding authors met in person to iteratively formulate categories of types of data communicated in clinical notes by merging similar codes from our final annotation scheme. These categories formed the taxa of a taxonomy of narrative note content. Note: at this preliminary stage, we avoided the use of generic types of data or taxa (e.g., other) and attempted to classify each annotated sentence in at least one type of data/taxon to avoid ambiguity.
The most common formats of HPI in primary care notes were story (notes = 36; mean size = 198 words, SD [131]), followed by list with categories (notes = 3; mean size = 99 words, SD [56]) and list without categories (notes = 1; mean size = 22 words, SD [0]). The most common formats of IP in primary care notes were list with categories (notes = 32; mean size = 288 words, SD [155]), followed by story (notes = 8; mean size = 197 words, SD [59]). The most common formats of HPI in specialty care notes were story (notes = 21; mean size = 130 words, SD [52]), followed by list without categories (notes = 14; mean size = 113 words, SD [55]) and list with categories (notes = 5; mean size = 92 words, SD [47]). The most common formats of IP in specialty care notes were list with categories (notes = 23; mean size = 123 words, SD [54]), followed by story (notes = 17; mean size = 233 words, SD [116]). None of the notes had the IP section classified as list without categories. Table 2 summarizes the average size of each note section per note format and specialty.
We identified five combinations of format of HPI and IP sections respectively: (A) story + list with categories (notes = 36; mean size = 415 words, SD [222]); (B) story + story (notes = 21; mean size = 396 words, SD [156]); (C) list without categories + list with categories (notes = 15; mean size = 243 words, SD [81]); (D) list with categories + list with categories (notes = 4; mean size = 312 words, SD [110]); and (E) list with categories + story (notes = 4; mean size = 425 words, SD [115]). There was a statistically significant difference between the combinations of formats as determined by one-way ANOVA (F(4,75) = 2.75, p = 0.03). A Tukey post-hoc test revealed that HPI and IP content was significantly smaller in combination C compared to combination A (-172 words, [95% Conf. -326, -17.89]; p = 0.02). Table 3 summarizes the size of each combination of narrative note format and Figure 2 illustrates the size and the number of notes by combination of formats.
After the definition of an initial coding scheme and an initial round with independent annotation of 15 notes, inter-rater agreement was 83%, K .83, we therefore split the remaining sample between the two coding authors.
We identified 45 types of data communicated in narrative note sections. The most common types of data in HPI in primary care notes were “history of intervention reported” (n = 106) and “history of finding/condition” reported (n = 106), followed by “description or value of a finding/condition” (n = 87). The most common types of data in IP in primary care notes were “intervention planned or recommended” (n = 299), followed by “date of an intervention” (n = 220), and “finding/condition present/presumed” (n = 114). The most common types of data in HPI in specialty care notes were “history of intervention reported” (n = 98), followed by “history of finding/condition reported” (n = 73), and “date of an intervention” (n = 65). The most common types of data in IP in specialty care notes were “intervention planned or recommended” (n = 110), followed by “intervention performed during the visit” (n = 64), and “reason for ordering or performing an intervention” (n = 43). We classified the types of data into seven taxa: 1 - finding/condition documented (n = 14), 2 - intervention documented (n = 9), 3 - general descriptions and definitions (n = 7), 4 - temporal information (n = 6), 5 - reasons and justifications (n = 4), 6 - participants and settings (n = 4), and 7 - clinical documentation (n = 1). Figure 3 illustrates the taxonomy of narrative note content and Table 4 lists the 45 types of data communicated in narrative note sections.
We systematically assessed and classified the most commonly used narrative note section formats and narrative content reported in primary and specialty care notes. Although assessments of clinicians’ preferences about note formats and about the expressivity of clinicians’ reasoning found in narrative note sections have been previously reported 24 , to our knowledge, no systematic assessment of the most common formats used in narrative note sections is available in the literature. The quantitative assessment of note content redundancy and the deleterious effects of “copy-and-paste” have been also reported 14 , but an assessment of the most common types of data communicated in narrative note sections has not. Our findings will empower the broader medical and informatics communities by facilitating the identification of note content that is more commonly found in narrative note sections, and the most commonly used narrative note section formats to help informatics researchers to tailor their efforts and design clinical documentation systems that can accommodate common structures for documenting patients’ care context data.
Primary care notes were found to have significantly longer narrative note sections when compared to specialty care notes. This is likely explained by the fact that primary care physicians tend to have a holistic view of the patient and usually treat or at least follow up on most patient’s present or presumed conditions, whereas specialty care clinicians tend to focus on specific findings and conditions that require immediate or long-term specialized care 25 . Although the overall note size was not significantly different between the two groups, primary care notes had significantly higher proportion of content in HPI and IP sections than specialty care notes. This is also consistent with a larger number of findings and conditions treated by primary care physicians. Our annotations demonstrate that the note section HPI was more frequently documented as a story in both groups, and the most common types of data in this section include “history of intervention reported” and “history of finding/condition reported,” but these are more frequent in primary care notes (106 instances vs. 98 instances and 106 instances vs. 73 instances, respectively), indicating a greater variety of concepts documented in primary care documents. A similar pattern was observed in the IP section with the most common format being list with categories in both groups, and the most commonly reported type of data in both groups being “intervention planned or recommended,” but with nearly triple the frequency of interventions found in primary care notes (299 instances) when compared specialty care notes (110 instances).
Not surprisingly, our annotations revealed that narrative note sections frequently contain contextual information about the patient’s health and care and clinicians’ decisions. The top five types of data in HPI in both groups represent 40% of the overall HPI annotations, and include documentation of past or present findings, conditions, and interventions, as well as the description or value of findings or conditions (e.g., hemoglobin/hematocrit: 8/26). The top five types of data in IP represent 45% of the overall IP annotations and include interventions planned or recommended by the clinician as well as contextual data about intervention such as their date (e.g., return to clinic in three months) and the reasons or justifications for ordering them (e.g., return to clinic if symptoms reoccur). However, another important element of patients’ care context data, the clinician’s interpretations, which we annotated as “clinical interpretation of findings/conditions or interventions,” was not found as frequently as the other components of patients’ care context.
Overall, clinicians’ interpretations were only the 16 th most common type of data found in our annotations. In HPI, it was found in only six instances in primary care notes and only once in specialty care notes, and most instances referred to interpretations likely expected to be found in the IP section (e.g., patient demonstrates good inhaler technique). In IP, it was slightly more frequent but ranked moderately low compared to other types of data (9 th in primary care notes and 10 th in specialty care notes). We found 57 instances of clinician’s interpretations in the IP section in primary care notes and 22 instances in specialty care notes.
We believe that several factors contribute to the relatively low frequency of documentation of clinicians’ reasoning. First, the steady implementation of pay-for-performance payment models increased the demand for capturing accurate, structured data 26 , which could force clinicians to more frequently document (or import) billing-associated data into their notes. This is reinforced by the fact that the need to document non-clinically relevant data in clinical notes is an increasing source of frustration among clinicians 11 . In our annotations, coded diagnoses were more frequent than clinical interpretations in IP in specialty care notes (24 instances vs. 22 instances). Second, the coupling of EHR limitations and strict coding/billing requirements are also considered a contributing factor for the longer notes found in the U.S. 12 , and could as well contribute to the documentation of non-clinically relevant information at the expense of patients’ care context data. Third, previous studies suggest that clinicians tend to make conscious decisions about what to document in their clinical notes 27 , and documentation or the lack thereof of interpretations could be one of these conscious decisions. Finally, previous studies suggest that current EHRs provide virtually no automation for documentation of clinicians’ reasoning in their clinical notes 24 , which provides opportunities for developing clinical documentation systems that can help clinicians’ more frequently document and access their own interpretations, as well as other clinicians’ interpretations about patients’ health and care.
Implications for future research and EHR development. The importance of communicating what clinicians are thinking about the patients and their problems has already been reported by Weed in 1968 28 , however, a half-century later we still lack key information about patients’ care context in their health records, which coupled with bloated digital notes, introduces data obfuscation and facilitates medical errors 5 . Further research is needed to explore clinicians’ perceptions about preferred narrative note section formats and relevant narrative content, since frequency of use does not necessarily assure usefulness of content. Our work can guide research focused on the development of clinical documentation systems that facilitate creation and retrieval of concise notes, especially for the creation and synthesis of narrative content relevant to patient care. This may require a combination of a formal representation of narrative note content 23 to augment prominent computational methods such as conversational speech recognition and natural language processing 3 . These methods could be used to automatically capture clinical and contextual data often verbalized during clinical visits 23,27 and, when combined with other modern language understanding methods 29 improve data entry at the point of care, by facilitating automatic creation of narrative note content that adheres to the preferred narrative formats and relevant content.
Limitations. Although our taxonomy was created with a robust sample of notes and annotations, its taxa include data from only one institution; notes from other institutions may reveal other types of data. We included only three clinical specialties; notes from other specialties may contain additional formats and data types. However, our methods are reproducible and can be relatively easily applied to other institutions and specialties. We did not assess other note sections such as “review of systems” or “health maintenance,” which in some cases may contain context data not captured by our annotations. However, we believe that these instances are rare and, in most cases, include information already documented in HPI and IP, therefore, our annotations may have covered most patients’ care context data documented by clinicians. Known influences to note creation such as copy-and-paste or the use of templates were out of the scope, but we believe they may influence creation of narrative content and warrant future studies.
We assessed and classified the most commonly used narrative note section formats and narrative content reported in primary and specialty care notes. Narrative note sections tend to be significantly longer in primary care than in specialty care notes, despite a similar overall note size between these groups. We identified commonly used narrative note section formats and developed a taxonomy of narrative note content to help researchers to tailor their efforts and design more efficient clinical documentation systems.
This work was supported by funds from the Informatics Institute of the University of Alabama at Birmingham.