Background
Coronary heart disease affects 18.2 million adult Americans and over 1.3 million yearly hospitalizations involve acute coronary syndrome [
1]. The transition from hospital to home can be challenging for patients. Patients with acute coronary syndromes often experience warning symptoms after hospital discharge, leading them back to seek acute care [
2‐
4]. The 30-day rehospitalization rate for adult patients (≥ 18 year of age) has ranged from 10 to 17% [
1,
5,
6].
Evaluation of patients’ post-discharge symptoms is an integral part of transitional care to improve clinical outcomes including readmission [
7]. Pain (both cardiac and non-cardiac) is common in post-discharge patients with acute coronary syndromes [
8‐
13]. Research evidence of the impact of pain on healthcare utilization is valuable to transitional care, but was mainly on cardiac pain for this patient population [
2,
3,
8‐
10]. Little is known about generic pain and its association with rehospitalization.
In this study, we evaluated patient-reported generic pain within 30-days post-discharge and its association with 30-day rehospitalization in patients with acute coronary syndromes, using data from a prospective cohort study, the Transitions, Risks, and Actions in Coronary Events Center for Outcomes Research and Education (TRACE-CORE) project [
14]. We studied generic, not just cardiac, pain because in a previous study we found that generic pain was associated with 30-day acute care use in patients with heart failure [
15]. Also, although cardiac pain is typical for patients with acute coronary syndromes, other types of pain (e.g., non-cardiac chest pain, wound pain from surgery, and chronic pains) are also common [
11‐
13,
16‐
18]. Recent studies found that non-cardiac or non-specific chest pain was prevalent in readmissions in patients with acute coronary syndromes [
11,
12,
19,
20]. Further, the cardiac nature of pain is not always apparent, as is captured in the nomenclature “atypical chest pain” and “non-specific chest pain” [
21,
22], which may refer to generic pain or other angina-equivalent symptoms. We hypothesized that post-discharge patient-reported generic pain would be associated with 30-day rehospitalization. In addition, we examined the associations among patient-reported pain, documentation of pain in electronic health record (EHR) notes, and rehospitalization. Mentions of pain in EHR notes were automatically identified by natural language processing (NLP).
Methods
Data sources
Data for this study were obtained from the TRACE-CORE cohort study [
14] and enhanced by linking the original TRACE-CORE data to patients’ electronic health record notes within the healthcare system of their index hospitalization (UMass Memorial Health Care—UMMHC in Worcester, MA).
Study design
In TRACE-CORE, patients hospitalized for acute coronary syndromes were followed prospectively through patient interviews at 1, 3, 6, 12 months post-discharge and medical record abstraction for 24 months [
14]. We examined pain reported by TRACE-CORE patients at one month post-discharge and its association with 30-day rehospitalization.
Setting and sample
The details of the study setting and patient recruitment have been documented for TRACE-CORE elsewhere [
14,
23]. Briefly, patients were recruited from 6 study sites (hospitals), which serve the majority of patients hospitalized with acute coronary syndromes in central Massachusetts and central Georgia, between April 2011 and May 2013 [
14,
23]. The cohort included adult (over 21 years old) patients who had at least one of: (1) serial ECG changes consistent with ischemia; (2) elevation of cardiac biomarkers; (3) cardiac catheterization revealing over 70% stenosis in a coronary artery; and (4) admission for urgent or rescue PCI/CABG, with symptoms of acute ischemia in the 72 h prior to admission. Exclusion criteria included pregnancy, prison custody, receiving palliative care, dementia, acute coronary syndrome secondary to demand ischemia or aortic dissection, perioperative acute coronary syndrome, admission for trauma or elective cardiac catheterization procedure, and transfer from another hospital after staying at the referring hospital for over 24 h.
For this study, we included the TRACE-CORE patients who were discharged from two hospitals that comprise UMMHC and also completed the assessment of generic pain at on-month follow-up (N = 787). We chose this dataset because we had access to EHR notes from UMMHC for the TRACE-CORE patients for additional analyses involving EHR-documented pain.
Institutional Review Board (IRB) approval was obtained for the parent TRACE-CORE study from the Committee for the Protection of Human Subjects in Research at the University of Massachusetts Medical School (UMMS) and the other study sites. All participants provided written informed consent. Additional IRB approval was obtained from the Committee for the Protection of Human Subjects in Research at UMMS to access EHR notes for this study.
Data collection
Baseline variables
Baseline data for each patient were collected through a computer-assisted 60-min structured interview during patient’s index hospitalization or by phone call within 72 h post-discharge, which included patient’s demographics, socioeconomic factors, healthcare seeking behavior, and other background information. Patient’s medical history, comorbidities, and cardiac procedures were obtained through medical record abstraction. Healthcare seeking behavior was assessed by two questions “Is there a place that you usually go to when you are sick or need advice about your health?” (with 3 response options: (1) yes, (2) there is no place, and (3) there is more than one place) and “What kind of place is it—a clinic, doctor’s office, emergency room, or some other place?” (with 4 response options (1) clinic, (2) doctor’s office, (3) hospital emergency room, and (4) some other place”). Because only patients who answered (1) “yes” to question 1 have responded to question 2 in our data, we combined responses to these two questions into a single 4-category variable (see Table
1). We combined response options (1) “clinic” and (2) “doctor’s office” for question 2 into one category that represented the non-acute care setting. This treatment was informed by prior studies in health services research that separated non-acute care and acute care when studying factors impacting hospital admissions and strategies to reduce admissions [
24‐
26]. Comorbidity was estimated by counting diseases from 14 medical conditions (see footnote in Table
1) documented in patient’s medical history. Table
1 names baseline variables included in this analysis.
Table 1
Patient characteristics at baseline, by pain self-reported one month after discharge: TRACE-CORE, 2011–2013
| N = 787 | N = 285 | N = 101 | N = 347 | N = 54 | |
| N (%) | N (%) | N (%) | N (%) | N (%) | |
Age | | | | | | 0.01* |
< 50 | 116 (14.7) | 30 (10.5) | 14 (13.9) | 64 (18.4) | 8 (14.8) | |
50–64 | 336 (42.7) | 120 (42.1) | 47 (46.5) | 141 (40.6) | 28 (51.9) | |
≥ 65 | 335 (42.6) | 135 (47.4) | 40 (39.6) | 142 (40.9) | 18 (33.3) | |
Gender | | | | | | 0.01* |
Female | 235 (29.9) | 68 (23.9) | 26 (25.7) | 121 (34.9) | 30 (37.0) | |
Race | | | | | | 0.42 |
White | 741 (94.9) | 269 (95.1) | 97 (97.0) | 326 (94.8) | 49 (90.7) | |
Black | 8 (1.0) | 4 (1.4) | 1 (1.0) | 3 (0.9) | 0 (0.0) | |
Hispanic | 32 (4.1) | 10 (3.5) | 2 (2.0) | 15 (4.4) | 5 (9.3) | |
Education | | | | | | 0.31 |
≤ High school | 286 (36.3) | 115 (40.4) | 32 (31.7) | 116 (33.4) | 23 (42.6) | |
Some college | 241 (30.6) | 75 (26.3) | 33 (32.7) | 116 (33.4) | 17 (31.5) | |
≥ College graduate | 260 (33.0) | 95 (33.3) | 36 (35.6) | 115 (33.1) | 14 (25.9) | |
Low social support | | | | | | 0.38 |
Yes | 41 (5.2) | 17 (6.0) | 5 (5.0) | 14 (4.0) | 5 (9.3) | |
Live alone | | | | | | 0.63 |
Yes | 158 (20.1) | 54 (18.9) | 24 (23.8) | 67 (19.3) | 13 (24.1) | |
Low health literacy | | | | | | 0.69 |
Yes | 227 (29.0) | 80 (28.2) | 34 (34.0) | 97 (28.0) | 16 (29.6) | |
Heavy drinker | | | | | | 0.30 |
Yes | 74 (9.4) | 25 (8.8) | 13 (12.9) | 34 (9.8) | 2 (3.7) | |
Current smoker | | | | | | 0.04* |
Yes | 161 (20.6) | 52 (18.2) | 23 (22.8) | 68 (19.6) | 19 (35.2) | |
Where to seek health care | | | | | | 0.44 |
No places | 87 (11.1) | 28 (10.0) | 11 (10.9) | 42 (12.1) | 6 (11.1) | |
Clinic or doctor office | 619 (79.2) | 233 (82.9) | 78 (77.2) | 267 (77.2) | 41 (75.9) | |
Other places | 35 (4.5) | 8 (2.8) | 4 (4.0) | 21 (6.1) | 2 (3.7) | |
Emergency department | 41 (5.2) | 12 (4.3) | 8 (7.9) | 16 (4.6) | 5 (9.3) | |
Acute coronary syndrome type | | | | | | 0.01* |
STEMI | 171 (21.7) | 72 (25.3) | 10 (9.9) | 79 (22.8) | 10 (18.5) | |
NSTEMI | 470 (59.7) | 174 (61.1) | 65 (64.4) | 200 (57.6) | 31 (57.4) | |
Unstable angina | 146 (18.6) | 39 (13.7) | 26 (25.7) | 68 (19.6) | 13 (24.1) | |
History of PCI | | | | | | 0.12 |
Yes | 174 (22.1) | 54 (18.9) | 22 (21.8) | 80 (23.1) | 18 (33.3) | |
History of CABG | | | | | | 0.33 |
Yes | 87 (11.1) | 26 (9.1) | 16 (15.8) | 39 (11.2) | 6 (11.1) | |
Comorbiditya (mean [SD]) | 2.0 [1.7] | 1.8 [1.6] | 2.3 [1.8] | 2.0 [1.7] | 2.8 [2.0] | < 0.001* |
Reporting generic pain | | | | | | < 0.001* |
No | 181 (23.1) | 107 (37.8) | 11 (10.9) | 58 (16.8) | 5 (9.3) | |
Mild to moderate | 447 (57.1) | 141 (49.8) | 65 (64.4) | 220 (63.8) | 21 (38.9) | |
Severe | 155 (19.8) | 35 (12.4) | 25 (24.8) | 67 (19.4) | 28 (51.9) | |
Patient-reported post-discharge pain status
Patient’s post-discharge health-related quality of life, including pain, was assessed by the structured phone interview at one month post-discharge. The generic pain status was assessed by “How much bodily pain have you had during the past 4 weeks?—none, very mild, mild, moderate, severe, very severe” through SF-36 [
27]. Based on sample size, we grouped responses with “very mild”, “mild”, and “moderate” into a single category mild-to-moderate and grouped “severe” and “very severe” into severe in our analysis. Chest pain was assessed by the Seattle Angina Questionnaire (SAQ), “Over the past 4 weeks, on average, how often have you had chest pain, chest tightness, or angina?—4 or more times per day, 1–3 times per day, 3 or more times per week but not every day, 1–2 times per week, less than once a week, None over the past 4 weeks” [
28]. We grouped the first five categories into a single category having-chest-pain in our analysis.
Pain status documented in post-discharge EHR notes
We used NLP to categorize whether patients had documentation of pain in clinical notes. NLP has been increasingly used to process EHR data, including extracting symptoms from clinical notes [
29]. Researchers have also used NLP-identified symptoms for health services research [
30‐
34].
For this study, we developed and validated an NLP system to identify pain symptoms from EHR notes [
35]. In brief, using a knowledge-driven method, we adapted a general-purpose clinical NLP system cTAKES [
36] to pain extraction. cTAKES extracts and maps medical terms to Unified Medical Language System concepts [
36]. To adapt cTAKES to the pain domain, we first compiled a list of terms related to pain (e.g., pain, tenderness, headache, angina, etc.) using input from domain experts and then converted the term list to regular expressions to match cTAKES-recognized terms. We evaluated the system using 200 in-patient and out-patient notes (11,917 sentences), with each sentence labeled by a physician as mentioning pain or not. An EHR note was labeled as documenting pain if it contained at least one sentence that mentioned pain. We measured the system’s performance by three metrics commonly used for evaluating NLP: precision, recall, and the F1 score. In our case, precision is the number of pain instances correctly identified by NLP divided by the number of pain instances identified by NLP; recall is the number of pain instances correctly identified by NLP divided by the number of pain instances identified or labeled by domain experts; F1 is the harmonic mean of precision and recall. A pain instance is a sentence (for sentence-level evaluation) or an EHR note (for note-level evaluation) that documented pain. The system achieved 0.80 precision, 0.91 recall, and 0.85 F1 for pain extraction at the sentence level and 0.89 precision, 0.98 recall, and 0.93 F1 at the EHR note level.
We applied the system to clinical notes created after a patient’s discharge from index hospitalization and before rehospitalization (if any) to identify pain symptoms. If a patient was not re-hospitalized before the 1-month follow-up, we sought pain documentation in the EHR for the entire 1-month period. When applying NLP, we used a rule-based method to identify and exclude the “History of Present Illness” sections, which often documented patients’ chest pain related to their index hospitalizations rather than pain occurred after hospital discharge. EHR notes created after a patient’s rehospitalization could also document pain. We excluded those notes from our analysis, because we wanted to use the NLP-identified EHR-documented pain as a conceptual proxy for whether there was any provider-patient engagement around the patient’s pain status before readmission (if any) during the 30-day post-discharge period, i.e., whether the pain was “addressed”.
Outcome variable
Patients’ self-reported rehospitalizations within 30 days post-discharge were confirmed by chart review in the study hospitals when possible.
Data analysis
We first compared baseline patient characteristics across four pain categories defined by patient-self-reported pain at 1-month follow-up and EHR-documented pain using chi-square test, Fisher’s exact test, Cuzick's Test for Trend [
37], or analysis of variance as appropriate. The four pain categories were: (1) self-reporting no pain; (2) self-reporting mild-to-moderate pain and having EHR-documented pain before readmission (if any); (3) reporting mild-to-moderate pain but having no EHR-documented pain before readmission (if any); and (4) reporting severe pain. We did not separate patients reporting severe pain base on whether they had EHR-documented pain because only 2 patients reported severe pain, had EHR-documented pain, and were readmitted. This sample size is too small for the secondary analysis described below.
We then reported descriptive statistics about patient-self-reported pain and NLP-identified EHR documentation of pain. In a post-study analysis, we reviewed cases where patients did not report generic pain in survey but had NLP-identified EHR documentation of pain, and categorized the reasons for mismatch.
In the primary analysis, we first assessed the incidence of the outcome (i.e., 30-day rehospitalization) across the levels of patient-reported generic pain status (not considering EHR documentation) and the unadjusted association between patient’s pain status and the outcome. We then used multivariable logistic regression to adjust for potential confounding by (1) patient characteristics collected in index hospitalization and (2) patient characteristics collected in index hospitalization plus patient-reported chest pain at 1-month follow-up. We identified potential confounders from statistical analysis (Table
1,
P < 0.05). We also adjusted for healthcare seeking behaviors for both models 2 and 3.
Further, we conducted a secondary analysis to assess the relationship between the outcome, patient-self-reported pain, and EHR documentation of pain. We used the variable that represents the four pain categories (defined at the beginning of this section) as the predictor variable, adjusting for covariates as same as defined for the primary analysis.
As a sensitivity analysis to further disentangle the effects of chest pain versus generic pain, we repeated the association analyses, limiting them to patients not self-reporting chest pain by SAQ (n = 530).
Statistical significance was defined as a two-tailed
p-value of < 0.05. Statistical analyses were programmed using STATA/IC 15.1 (StataCorp LLC, Texas, USA) [
38].
Discussion
We found that after discharge from a hospitalization for acute coronary syndrome, patient-reported generic pain was prevalent and was associated with rehospitalization. After adjusting for patient-reported chest pain and other covariates, severe generic pain was still associated with rehospitalization. Going beyond our prior work that studied the association of pain and readmission [
15], by using the TRACE-CORE data and NLP, we were able to explore patient-provider engagement around pain, i.e., documentation of pain (presumably addressed) identified by NLP from clinical notes, within 30 days post-discharge. Notably, when pain was both self-reported and documented in the EHR, the rehospitalization rate was not significantly higher than that for patients reporting no pain (8.9% vs. 7.4%,
P = 0.62). However, the rehospitalization rate when pain was self-reported but NOT documented in the EHR was significantly higher than no pain (13.3% vs. 7.4%,
P = 0.02). Below, we further explore our findings by highlighting the unique aspects and discuss their implications.
We found a “dose–response” increase of 30-day rehospitalization rates across post-discharge pain severity levels, from no pain to mild/moderate pain to severe pain (Table
3), which persisted after adjusting for clinical factors collected from the index hospitalization. These results were compatible with our previous finding that patient-reported post-discharge generic pain was associated with increased risk of 30-day acute care use [
15]. Our prior results were in heart failure patients who reported post-discharge pain through daily automated telephone assessments [
40]. Taken together, both studies suggest that monitoring patient’s pain status after discharge may be valuable in identifying high-risk patients for early intervention to improve patient outcomes including readmission. However, prior studies do not address the questions: What happens if the patient’s pain is addressed by a healthcare provider? How will this impact readmission?
Using NLP methods, we examined presumed patient-provider engagement around pain before patient’s rehospitalization and assessed its impact on rehospitalization (Table
4). NLP revealed that, among patients self-reporting post-discharge pain, less than 20% had documentation of pain in their clinical notes before readmission, suggesting a lack of early attention to patients’ pain symptoms by providers. Even after adjusting for multiple covariates from the index hospitalization, patients with mild-to-moderate pain NOT documented in the EHR had more than twice the odds of readmission than no pain (OR: 2.03, 95% CI: 1.14–3.62). In contrast, the odds ratio for mild-to-moderate pain documented in the EHR versus no pain was not significantly different from one (OR: 1.23, 95% CI: 0.52–2.90). Thus, our findings are consistent with a clinical scenario where, for patients with pain, having pain attended by healthcare providers improves patients’ ability to stay at home to reduce early readmissions.
It is worth noting that our study used patient-self-reported pain collected by the 1-month follow-up survey (SF-36), which queried patient’s pain status during the past four weeks. We therefore cannot completely confirm that patient-reported pain always began before readmission. It is possible that patient-reports of pain began before readmission. This is compatible with our analysis results and the scenario discussed in the previous paragraph (Table
4). In addition, patients who reported severe pain had more EHR-documentation of pain before readmission compared with patients who reported mild-to-moderate pain (Table
2). This is consistent with a scenario where patients who had severe pain were more likely to seek care before readmission than those with mild-to-moderate pain. If the majority of patient-reports of pain began after readmission, it would be less likely to see these patterns from our data. To be sure, it is also possible that some patients were first readmitted to the hospital and then had new pain. For these patients, the new pain could be related to treatment received during rehospitalization or to other diseases. Future studies that include daily assessments of pain will help us better understand the directionality of the association between pain and readmission.
In studying post-discharge care and patient reports of pain, researchers and clinicians should take care to distinguish between cardiac and non-cardiac pain, especially in patients with acute coronary syndromes. In this study, we tried to separate the effects of cardiac and non-cardiac pains on readmission by adjusting for patient-reported chest pain through the SAQ. After adjusting for this factor (in addition to other covariates), the association between severe pain and 30-day rehospitalization remained significant; while mild-to-moderate pain (Model 3 in Table
3) and unaddressed mild-to-moderate pain (Model 3 in Table
4) were not significant any more. Thus, patient-reported chest pain is a partial confounder of the association between generic pain and readmission. Our sensitivity analyses, using only patients without self-reported chest pain—a much smaller sample, provided point estimates similar to the main analyses, although not statistically significant. This suggest that generic pain may, indeed, be an independent factor associated with rehospitalization. Note that separating effects of cardiac and non-cardiac pain is challenging in the real-world settings. First, multimorbidity is prevalent in older patients with cardiovascular disease [
41]. In our sample, patient-reported chest pain (SAQ) and generic pain (SF-36) co-occur frequently. Second, there is a measurement challenge. SAQ asks patients to report chest pain but cannot fully distinguish between cardiac and non-cardiac pain. Non-cardiac chest pain is common in patients with acute coronary syndromes [
11,
12], and it is difficult for patients to distinguish between cardiac and non-cardiac chest pain [
42,
43]. It is possible that patients would report both cardiac and non-cardiac pain (e.g., chest pains with musculoskeletal, gastrointestinal, or respiratory origin [
22,
44,
45]) when responding to SAQ. From the analytic perspective, more patients with generic pain and no reports of cardiac pain would be helpful in isolating the association of interest. From the perspective of patient care, the measurement challenge and the literature documenting anxiety, fear, physical disability, and high healthcare utilization in patients with non-cardiac chest pain [
43,
46‐
48] suggest that early evaluation and treatment of post-discharge chest pain may enhance patient safety and quality of life and also reduce excessive healthcare cost.
By using NLP, we were able to quantify EHR documentation of pain for the large sample of patients in this study. However, like other artificial intelligence technologies, NLP can make mistakes. As a means of quality control, we validated our NLP system on 200 clinical notes, which showed decent results including a high recall (see Data Collection), before applying it to this study. Our post-study analysis suggests that the NLP system needs further improvement in detecting negation, which we will address in future.
Our study has clear limitations. We used observational data, which, of course, precludes definitive causal inferences. Also, our data lack detailed information on the nature of post-discharge pian, such as onset, location, and duration. The 1-month follow-up survey queried patient’s pain status during the past four weeks, we therefore cannot ascertain every patient-reported pain occurred before readmission. Importantly, we used EHR documentation of pain as a proxy for the pain having been addressed by a clinician, which may not have occurred; and the lack of EHR documentation of pain may mean lack of care overall. To partially address this last possibility, we used patient-self-reported healthcare seeking behavior to adjust the analysis. We hope that our findings from this exploratory study will serve as a catalyst for future in-depth research about post-discharge non-specific pain in patients with acute coronary syndromes.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.