Capturing Clinician Reasoning in Electronic Health Records: An Exploratory Study of Under-Treated Essential Hypertension

Abstract

Monitoring response to antihypertensive medications is a frequent reason for outpatient visits. Blood pressure (BP) is often documented as elevated, but no change in medication occurs (Medication Non-adjustment or MNA). We studied the frequency of MNA, reasons for non-adjustment, how reasons (including reasons for patient nonadherence) were documented, and whether they could be represented in a clinical care context ontology. We examined 129 visit notes with MNA occurring in 80 cases (59%). We coded MNA as Conscious Maintenance (patient adherent but clinician continues therapy for stated reason), Nonadherence (clinician attributes BP elevation to patient nonadherence), and Finding Not Addressed (clinician does not indicate reasoning for MNA). We characterized Conscious Maintenance with 11 subcodes and Nonadherence with 6 subcodes. Our ontology successfully represented relationships between concepts and reasoning, supporting the feasibility of formal representation of clinical care contexts for patient care, decision support and research.

Introduction

"Your patient's blood pressure is elevated today. Are you going to adjust his antihypertensive medication?"

"No. He forgot to take his pills this morning."

In our experience, this exchange occurs countless times daily in clinics and medical practices, resulting in "medication non-adjustment" (MNA). Its occurrence indicates that the clinical encounter may be wasted, a loss for all involved.* Other reasons include running out of medications, a change in insurance coverage, confusion about what medications to take, etc., each of which could be addressed without an in-office blood pressure measurement. Of course, even when the patient has not adhered to medications, there may be value to the visit. However, in our experience there are many occurrences where the sole purpose of the visit is to assess medication effectiveness. We wondered about the frequency of such wasted visits, as well as other types of MNA visits, and naturally turned to our clinical data repository of electronic health record (EHR) data to answer the question:

1. How often does a visit in which an elevated blood pressure (BP) is detected result in an MNA visit?

   Information on diagnoses, vital signs and medications are all readily available in structured, coded form. However, the reasons for therapeutic decisions (including inaction) - the "why" in clinical reasoning - is not similarly represented.[1] For information such as a differential diagnosis or reason for choosing a particular therapy we need to review clinical documentation such as visit notes. This is a common task in clinical research: extracting findings and outcomes from narrative text, typically carried out manually, sometimes with the assistance of natural language processing (NLP).[2] This led to our second research question:

2. How does the clinician document the reason for MNA despite an elevated BP?

   We have recently begun developing an ontology for representing the context of clinical elements in EHR records that can be used to formally capture the "why" in the clinicians' notes, to make information that is currently buried in clinical notes interpretable by a computer. We have based our work on an analysis of the OBO Foundry,[3] surveys of the published medical literature on clinicians notes in EHRs[4] and clinical decision support systems,[5,6] and our own analyses of case reports, clinical notes and spoken communications during clinical visits.[7] Trying to understand the frequency of various types of MNA visits (a research question in its own right) presented an opportunity to conduct informatics research that attempts to answer two additional questions:

3. Does our ontology provide a means for representing the reason for the clinician's MNA decision?

4. Does the narrative text seem amenable to NLP in order to automate extraction of the reasoning?

We sought to answer these questions through an analysis of patient records. This paper provides preliminary results of that research.

Methods

Data Set

We searched our local i2b2 database for patients age 18 or greater with a diagnosis of essential hypertension (based on ICD10-CM "I10" codes) with one or more visits to ambulatory clinics in 2018. We further restricted visits to medical or cardiology clinics (including general/internal medicine, primary care, family medicine, cardiology and heart failure) in which a systolic BP greater than 140mmHg and a diastolic BP of greater than 95mmHg were recorded during the visit, and a clinic note was available. Patient visits that had no prior antihypertensive medications were considered out of scope and excluded from the study. Remaining visits constitute "High BP" visits. All data (vital signs, medication orders and visit notes) were manually de-identified. The project was approved by the Institutional Review Board of the University of Alabama at Birmingham (Protocol #300003240).

Extraction of Clinical Context from Visit Notes

The note text files were read to determine (1) medication adherence, (2) whether an antihypertensive medication was prescribed, (3) if a change in antihypertensive medication was prescribed (not a refill) including a different antihypertensive medication or a change in dose of the same antihypertensive medication, and (4) the reason ("why") for changing (or not changing) an antihypertensive medication type or dose.

High BP visits were first coded into the Clinician Action categories of "Medication Adjustment" (MA) or "Medication Non-Adjustment" (MNA) based on medication orders placed at the time of visit. For each MNA visit, we examined the entire visit note to determine whether the patient's adherence to antihypertensive medications was explicitly documented. If nonadherence was explicitly documented the visit was coded as a "Nonadherence" visit. Visits in which adherence was explicitly documented or not mentioned (that is, nonadherence was not explicitly documented), were coded as "Conscious Maintenance" visits if a reason for MNA was documented and "Finding Not Addressed" visits if no reason was provided. The visits were subcoded to characterize the types of reasons for Nonadherence or Conscious Maintenance. Figure 1 depicts this coding strategy.

Figure 1: Algorithm for selection of High Blood Pressure Visits, classification as Medication Adjustment or Medication Non-Adjustment (MNA) Visits, and coding of MNA Visits based on patient adherence to previously prescribed medications and reason documented by the clinician for not adjusting the patient's medication.

Two of the authors (JJC and HDM) met periodically throughout the visit coding process to discuss new findings and come to a consensus on how the visits should be coded and how reasons for Nonadherence or Conscious Maintenance codes should be classified into subcodes, which were defined according to the reasons documented by the clinician. When no explanation was given in the text file for the clinician's MNA decision, the subcode "Unknown" was assigned. Visits were selected at random after fitting the inclusion criteria described above. Coding was discontinued after reviewing 20 consecutive visits with no new subcode findings.

Ontology of Clinical Context

A sequence of research projects has led us to develop a "starter set" ontology for representing how clinical concepts in the patient record (such as findings, conditions, procedures and the patient herself) relate to each other with respect to the patient's clinical state.[4-7] Ultimately, we annotated 626 sentences from in-patient consult notes, outpatient visit notes, and transcripts of voice recordings from doctor-patient visits. This resulted in 48 unique relationships that related concepts of 14 categories (e.g., patient, clinician, finding/condition, and intervention) to each other. Additional concepts were created to characterize contextual information, such as a finding/condition being a side effect of an intervention, or a justification being a reason for an intervention. The formalism used in our annotations represents data as subject-predicate-object or concept-relationship-concept tuples such as patient-hasfinding/condition-hypertension, hypertension-hasintervention-amlodipine, amlodipine-hasreason-patient has failed hydrochlorothiazide. The ontology provides restrictions on relationships allowed to relate each pair of subject and object concepts. Two of the authors (JJC and TKC) reviewed the results of the text classification to match text mentions in the reviewed visit notes to the categories of subject and object concepts in our ontology and then reviewed the allowed relationships between the concept pairs, and selected, through consensus, the relationships that most closely represented the meaning of the original text mentions analyzed.

Results

Data Set

The query of our clinical data repository identified 51,978 patients with a diagnosis of essential hypertension, an encounter in 2018, and age 18 or greater. Of these, 38,158 had one or more visits to one or more of the 28 clinics of interest, with a total of 162,309 visits. Of these, 5,194 patients had a total of 6,987 High BP visits based on our criteria.

Our review process was terminated after a total of 129 visit notes for 96 unique patients were reviewed. Of these, 80 notes (62%) on 64 unique patients represented MNA visits. A total of 50 unique antihypertensive prescriptions were found (including various doses and combinations of 23 unique ingredients). There were 11 subcodes (59 cases) in which the patient took the prescribed medication and the clinician provided a reason for maintaining the same prescription despite an elevated BP (Conscious Maintenance). There were 6 subcodes (16 cases) in which a reason for the patient not taking the prescribed medication (Nonadherence) was documented. In 14 cases, the elevated BP was noted but no reason was given for MNA, and in 8 cases the BP elevation was not mentioned in the visit note. Figure 2 shows the coding of the visits and Table 1 show details of the subcoding.

Figure 2: Coding of visit notes according to the protocol in Figure 1. Note that some visits were coded with multiple reasons for MNA. Details of the subcodes are shown in Table 1.

Table 1: Subcodes of reasons documented in 80 patient notes by clinicians for not changing a patient’s BP medication, despite an elevated reading.

Main Code	Subcode*	Count	Example	Description
Conscious Maintenance	Will Monitor	14	He will keep a home BP log and send to me in 1 week	MD wants patient to monitor BP at home and report findings
	Objective evidence of normal tension	11	Blood pressure has been well controlled when taken by the patient at home.	Patients home bp recordings are lower that in MDs office
	Lower Repeat BP	9	In office blood pressure is automatic 136/74 (recheck), manual recheck L-122/82, R-120/80.	BP decreased on repeat measurement; notes indicate this as justification for not changing medication
	Proximate Stress	7	She thinks it is high today due to pain	BP increase attributed to stress, including pain and white coat hypertension
	Proximate Diet	6	He ate sausage this AM, so his BP is a bit high.	BP increase attributed to diet
	Mild 4 Hypertension		Benign HTN with CKD stage IV: BP fairly well controlled. BP close to target
	Patient Education	3	Discussed bp management. Options and alternatives were discussed.	Discussed actions to reduce BP (e.g., exercise, weight loss, smoking cessation, decreasing alcohol intake, dietary changes)
	Treatment Delayed	2	She was sent down to the emergency room by patient escort and nursing personnel from this office.	Treatment was delayed for various reasons including hospitalization or positive ilicit drug screening
	Partially 1 Improved		Improved, continue with current meds	BP noted by improved since last visit
	Consult	1	Will discuss with his cardiologist and see if we can increase coreg vs add amlo	Decision deferred to discussion with consultant
	Proximate 1 Medication		Elevated today (he is currently on steroid pack)	BP increase attributed to another medication
Nonadherence	Unknown	6	Off BP meds, drinks 4 cups coffee everyday, does not exercise, does not add additional salt. restart meds	No reason for nonadherence documented
	No Medication Available	4	Ran out of all of his medications a while ago so has not been taking any of them. Poorly controlled today but not taking his medications.	Patient does not have medication (might be many reasons for this)
	Memory Loss	2	Off the first two meds for unclear reasons. Only taking 1 med, not sure about the other, Family worried about memory, forgetting meds	Poor compliance due to memory disturbance
	Emotional Stress	2	She admits she has not been compliant with her medications for the past few months due to the grief of her mothers passing.	Poor compliance due to emotional stress
	Symptom or Sign attributed to Medication	1	Has reduced and come off BP medication due to muscle cramps.	Anti-hypertensive medication causingn side effects
	Self-Medication Management	1	It sounds like he takes his medicines however he basically takes them as he sees fit.	Patient does not adhere to medication schedule
Finding Not Addressed	Finding Not Addressed	22	Here today with persistent nausea, abdominal pain and just feeling so sick.	Elevated BP not mentioned in visit note

^*Some visits were coded with multiple subcodes.

Abbreviations: BP: blood pressure; MD: doctor of medicine; CKD: chronic kidney disease; IV: intravenous; PHI: protected health information.

Applying an Ontology of Clinical Reasoning

Our analysis of the sentences that discuss patient's high BP or hypertension pharmacotherapy revealed that the concepts and relationships from our ontology[7] can easily be applied to formally represent most concepts included in these sentences. In all cases, the patient has the diagnosis of hypertension (patient-hasfinding/condition), which has a recommended/ordered treatment (finding/condition-has_intention_of_intervention). For Conscious Maintenance cases, the intervention is being followed by the patient (patient or finding/condition-has_intervention) and the high BP has an explanatory reason (finding/condition is_caused_by). For Nonadherence cases, the intervention is not being followed by the patient (patient-hasabsenceofintervention), which has an explanatory reason (intervention [absent]-has_reason). For the cases in which high BP was not accessed, neither treatments nor the reasons are specified, thus they cannot be formally represented. Figure 3 illustrates the formal representation of each case using the concepts and relationships from our ontology.[7]

Figure 3: Representation of major codes for documentation of medication non-adherence visits using an emerging ontology for formal representation of clinical care contexts[7]: (A) Conscious Maintenance (clinician acknowledges the BP elevation and medication adherence by the patient but continues current therapy for some stated reasons), (B) Nonadherence (clinician acknowledges the BP elevation and attributes it to the patient not taking medication, with or without documentation of the patient's reason), and (C) Finding Not Addressed (clinician does not acknowledges the BP elevation).

Discussion

In our limited sample of clinic visits, 62% were found to be MNA visits.2 This number would have been much higher if our selection criteria for BP visits had considered elevation of either systolic or diastolic BP to be sufficient for inclusion, rather than requiring elevation of both. We studied 80 visit notes of patients with BP elevation that was not followed by a change in hypertension pharmacotherapy. We found that MNA visits related to essential hypertension occur frequently and that reasons of non-adjustment are usually documented in visit notes. We also found that while the majority of cases involved maintaining existing medication for some secondary consideration, a significant portion of MNA visits involved non-adherence to prescribed medications due to a variety of causes. Our study has the limitations inherent in using a data set focused on a single condition from a single institution. Although confirmation in other settings is necessary, we have at least disproven the null hypothesis that clinicians do not document reasons or justifications for non-adjustment in their assessments and plans.

The ontology used in this study builds on prior work on formal representation of clinical context [6,7] which was required in the absence of the existence of appropriate terminologies in the OBO and the larger Unified Medical Language System (UMLS).[4] The subcodes in this study were developed for use with hypertension-related primary care visit notes. While some are fairly domain-specific( e.g., "Lower Repeat BP"), many should be relevant to any situation (disease and setting) in which a clinician is considering medication adjustment (e.g., "Memory Loss"). As with any extension of knowledge representation into a new domain, additions will likely be needed. Once more general applicability is demonstrated, our ontology should be suitable for inclusion in open source ontology repositories such as OBO and UMLS.

While our study can be considered preliminary and cannot be used to estimate the overall impact of subjective factors such as patient forgetfulness, the results are suggestive of a potential target for improving the efficiency of routine visits, acute visits, and patient care in general. Our findings can inform future studies that would benefit from being able to access the clinician's documentation for why the patient did not take their medication and why the clinician did not take any action. Such studies are greatly facilitated by the current availability of a digital infrastructure produced by nationwide adoption of EHRs;[8] however, while this infrastructure facilitates extraction of structured data (e.g., diagnosis of hypertension and orders for antihypertensive medications), methods for extracting information from narrative text (which includes reasons for not changing a medication) are often suboptimal.[9] Therefore, use of structured data, where available, continues to be the preferred method for reuse of EHR data in research[10], and the need for improving computational methods capable of processing narrative data has become paramount.[11]

EHR data are also a fundamental resource for the logic used in clinical alerting systems. These systems have traditionally been most successful using structured data, with narrative text generally being inaccessible for processing alert logic. For example (and coincidentally relevant to the subject of MNA visits), in a study of the impact of a commercial EHR implementation at a large, integrated care delivery network, Colicchio and colleagues found that an alert recommending treatment of patients with elevated BP could not be overridden, resulting in the documentation of high BP for cases in which the clinician generally considered the patients to be "in control" (Conscious Maintenance).[12] When interviewed, the clinicians suggested that when using these data for research, changes to antihypertensive treatment should be considered as a covariate for a "true" assessment of patients "not in control."

Reuse of EHR data for research and decision support (as well as billing, quality assurance, administrative functions, etc.) notwithstanding, our examination of the scientific literature shows, not surprisingly, that the primary reason for clinicians to document patients' findings, and their own interpretations and decisions is for patient care, while the strategies for retrieving and synthesizing such information vary with care goals.[6] One of the current challenges in EHR functionality is navigation of clinical documentation in the face of information overload.[13]

We believe that formal representation of the clinician's observations, interpretations, questions and decisions - what we refer to as the patient's care context - can advance informatics and clinical research, decision support, learning health systems, and information navigation. Consider the possibilities if the clinician's documentation of MNA visits were well-structured. A health outcomes researcher would be better able to tell when a medication was working or not working by knowing when patients are compliant or noncompliant and the reasons for the latter (side effects, cost, cure, forgetfulness, etc.). A clinical decision support system could avoid inappropriate alerts (by being able to process in its logic such contextual elements), thus reducing the need for clinician override and its attendant alert fatigue. And a clinician, wishing to determine next steps in a patient's care plan could be provided with a succinct summary of past medications and assessments of their effectiveness, which could facilitate navigation and information synthesis. The added knowledge could also be useful for a learning health system by adding missing contextual elements to the analysis of large-scale data sets.[14]

There are, of course, challenges to obtaining structured clinical context data. We used a manual approach for this study to assess the availability of information in the EHR, but we do not expect that manual extraction will be reproducible or even feasible on a large scale. Extracting clinician reasoning for MNA and other decisions for retrospective analysis (for research and quality assurance) and in real time (for decision support and improved clinical workflow) will need to rely on automated methods.

We are encouraged by our findings that (a) the rationale for a clinician's decision is generally present in visit notes, (b) unknown (not justified/addressed) cases seem to be relatively rare, and (c) the rational, at least at a generic level, seems to involve a fairly small number of concepts. We were also encouraged that the ontology we are developing is capable of representing these concepts and their relationships to patient findings and conditions and to each other. However, the wide variety of phrasing used to represent the reasoning is so broad (with subcodes rarely being expressed more than once in the same exact way) suggests that the use of NLP to extract context will be even more difficult than inherently simpler domains like patient symptoms.[8]

If NLP is not feasible, we will need to consider other methods by which clinicians can document context in structured form, which may likely emerge from the combination of prominent methods such as conversational speech recognition and other modern language understanding methods. [15,16] Given the current concerns about documentation burden as factor in clinician burnout, solutions will be challenging.[17] However, structured documentation in EHRs has been successful for certain tasks when there are tangible benefits. The benefits of structured documentation of clinical context (including the ability to check for completeness) will need to be made similarly apparent to the clinicians actually tasked with doing the documentation.

Conclusion

We have explored the documentation of patients' adherence to antihypertensive medication prescriptions and clinician's interpretations and reasoning for choosing not to adjust antihypertensive medications, when patients are and are not adherence to prescribed therapy. We found that documentation does occur and can be represented in our evolving ontology of clinical context. Both of these findings are necessary, although not sufficient, for moving forward with developing ways to formally capture such information to support multiple uses of EHR data, including patient care, decision support, and clinical research.