Uncovering hidden trends: identifying time trajectories in risk factors documented in clinical notes and predicting hospitalizations and emergency department visits during home health care

INTRODUCTION

Home health care (HHC) offers personalized healthcare, such as nursing and social support services, to patients in their homes.¹ In the United States, HHC is one of the fastest-growing healthcare sectors due to a rapidly aging population and the need to accommodate an individual’s desire for alternatives to institutional care.^2,3 HHC aims to assist patients in recovering from illness or injury, managing chronic conditions, maintaining independence, and minimizing the need for acute care services.⁴ Despite ongoing efforts to reduce negative outcomes, over 20% of patients experienced hospitalizations and emergency department (ED) visits during HHC services.⁵ Up to 40% of these negative outcomes are preventable with timely care^6–8; early identification of HHC patients at risk can lead to closer surveillance and earlier interventions to prevent hospitalizations or ED visits.⁹

Previous studies have used standardized assessments and other structured data to identify risk factors associated with hospitalizations or ED visits in HHC.^10–14 However, a significant portion of HHC risk information is not solely captured in structured data or standard assessments but is also found in unstructured data such as clinical notes.¹⁵ Prior research has shown that utilizing natural language processing (NLP) to analyze clinical notes has allowed the extraction of additional risk factors for hospitalizations or ED visits in HHC compared with studying structured data alone.^16–18 Our team found that including information from HHC clinical notes through NLP can significantly improve the ability of machine learning algorithms to predict the patients at risk for hospitalizations or ED visits.¹⁹

To identify risk factors in clinical notes, it is imperative to use standardized terminology which can serve as a reliable guide to risk identification in clinical notes. Standardized terminology promotes consistent and clear identification and documentation of risk factors, which in turn facilitates evidence-based treatment and prevention strategies.^20,21 The Omaha System comprises 3 key components: the Problem Classification Scheme, the Intervention Scheme, and the Problem Rating Scale for Outcomes. These components enable healthcare professionals to precisely document and monitor patient progress. Furthermore, the Problem Classification Scheme—which guides the identification of risk factors throughout this study—includes categories for health status, environmental factors, and health-related behaviors, resulting in a comprehensive understanding of the patient’s condition.^22,23 The Omaha System was developed to represent aspects unique to community-based care. In our previous NLP study, we used the Omaha System to identify risk factors for hospitalization or ED visits in HHC.^16,19,24,25

In HHC, clinicians conduct multiple home visits to patients throughout episodes of care (ie, all services provided between the patient’s admission and discharge from the HHC or within 60 days of the recertifying period), enabling longitudinal information to be collected. Longitudinal data, also known as time series data include multiple measurements, observations, and clinical notes documented at different intervals during the patient’s care, providing information about their clinical condition over time and allowing for analysis of trends and patterns. However, one limitation of previous HHC risk identification studies is that NLP-extracted risk factors were aggregated and analyzed at the episode level.^13,17,19,26 As a result of this aggregation, changes in risk factors over time were not examined. Clustering techniques can aid in analyzing time trajectories of risk factors by grouping similar patterns together, enabling the identification of representative patterns in time series data.²⁷ Despite the potential benefits of clustering for understanding and analyzing time series data, this approach has not been applied in HHC.

To address limitations in prior research, the aims of this study were to: (1) identify the clusters of temporal risk patterns documented in HHC clinical notes, (2) examine the association between the clustering in temporal risk patterns and hospitalizations and ED visits, and (3) identify the inter-relationships between the risk factor temporal domains.

MATERIALS AND METHODS

We conducted a retrospective cohort study utilizing 2 data sources: (1) structured data, consisting of the Outcome and Assessment Information Set (OASIS) and other assessment items extracted from the electronic health record (EHR) and (2) unstructured data (ie, clinical notes). This study was approved by the Institutional Review Boards of the participating institutions.

RESULTS

During the study period, 8227/73 350 (11.2%) of HHC episodes resulted in a hospitalization or ED visit. The average patient’s age was 79 years, and 64.3% were female. Among the sample, 63% of the patients were non-Hispanic Whites patients, followed by 17.2% of non-Hispanic Blacks patients, 13.4% of Hispanics patients, and 6.1% of others. For patients experiencing hospitalization or ED visits, the time gaps between the last note used for clustering and the hospital/ED visits varied. The minimum time gap was 0 days, showing that some hospitalizations or ED visits happened on the same day as the home visit. The median, mean, and maximum time gaps were 3, 11.1, and 59 days, respectively.

DISCUSSION

The study was the first to have identified the temporal patterns of risk factors based on the Omaha System documented in HHC clinical notes and examined the relationship between the temporal patterns and the risk for hospitalizations and ED visits. The analysis revealed 6 clusters that exhibited distinct patterns in how risk factors were documented over time within the Omaha System Problem Classification Scheme. Further, different patterns of time trajectory of risk factors showed varying effects on hospitalizations and ED visits.

Given the nature of HHC services, there are challenges integrating time-related factors into clinical risk assessment analyses due to the variations in visit frequency, length of stay, and different reasons for early discharges, such as recovery from the disease or deterioration that requires hospitalization or an ED visit. Because of these complexities, traditional methods (eg, latent class analysis) and general time series analysis are insufficient for analyzing HHC visit-level data because these methods assume equal number of data points (ie, visit frequency) and/or equal time intervals among all patients.^39–41 To address these challenges, this study implemented a unique and innovative dynamic time warping method that can handle nonlinear temporal patterns and time series with variations with minimal loss of time-related information.³² We also used unsupervised clustering methods to identify 6 temporal trends patterns in risk factors documented in clinical notes over time. These patterns exhibited positive or negative directional trends and a trend line with varying degrees of the slope or fluctuations. As opposed to linear or logistic regression,^42,43 cluster analysis captured slope fluctuations, for example Cluster 5, which initially had a decrease in documented risk factors, but the number of risk factors rebounded. This innovative method, which employs temporal cluster features, offers a valuable strategy for analyzing not only HHC visit data but also measurement data in hospital EHR (eg, vital signs). This is particularly useful as measurements are often taken at varying times for each patient with unequal time intervals, which can make analysis challenging.

In this study, we examined the clinical characteristics of different clusters in the cohort. While Cluster 3 “Decreased,” which has a majority of non-Hispanic White patients, showed a decrease in documented risk factors over time (64.7%), all other ethnicities were more prevalent in Cluster 1“No Risk Factors.” These results indicate that ethnicity may play a role in disparities in healthcare, as certain ethnic groups may have distinct health challenges or higher rates of specific health conditions,^44–47 which may be attributed to earlier or more aggressive strategies in addressing the risk factors specific to certain ethnicities. An alternative explanation might be that patients other than White patients are less likely to have their risk factors documented, despite reporting more problems during verbal communication with healthcare providers.⁴⁸ Our results also showed high prevalence of patients living alone in Cluster 6 “Steadily Present” compared with Cluster 1 “No Risk Factors.” This suggests that caregiver support is important for maintaining symptom control and preventing negative outcomes.⁴⁹

While there were statistically significant differences in other demographic and clinical factors extracted from structured data including OASIS, these differences between the clusters were relatively small; thus, their impact was not considered clinically relevant. The nature of the data may explain these relatively small differences: assessment data for OASIS start-of-care documentation are captured primarily during the first visit, while risk factors extracted from clinical notes are reported over time in multiple visits. Consequently, differences may exist between the characteristics of the cohort observed in the time trajectories of risk factors derived from clinical notes, which include the time aspect, and those assessed through the OASIS system. As a result, standard assessments at one time period cannot provide a complete representation of the changes that occur during a patient’s time in HHC. Therefore, further research is required to explore ways to capture changes in patient characteristics, along with temporal clustering of structured and unstructured data.

Our study examined the association between clusters and risk for hospitalizations or ED visits. Notably, compared with Cluster 1 “No Risk Factors,” the risk of hospitalization or ED visits was higher in the other clusters, due to the documentation of risk factors regardless of the directional trends or degree of slope Cluster 4 “Steeply Increased,” showed the highest risk of hospitalization and ED visits (OR, 2.95), followed by Cluster 5 “Decreased and Rebound Increased” (OR, 2.47). As seen in Cluster 5 “Decreased and Rebound Increased,” if new risk factors emerged or if patients experienced a cycle of illness, recovery, and then a recurrence of symptoms, the risk of hospitalization or ED visits may be higher compared with Cluster 6 “Steadily Present.” Furthermore, compared with Cluster 2 “Steeply Decreased,” Cluster 4 “Steeply Increased” demonstrated higher odds of hospitalization or ED visits. Based on these findings, it is reasonable to suggest that time-related variables are significantly associated with risk of hospitalization or ED visits. However, the method used in this study (ie, a dynamic time-warping algorithm to identify patterns based on similarities and an unsupervised hierarchical clustering) may not be able to distinguish between the differences in the number of risk factors, which could result in some patients with different numbers of risk factors being clustered together. Therefore, further research is needed to address this limitation and better understand the associations between risk factors and hospitalization or ED visits. These findings also suggest that the risk factors for hospitalizations or ED visits identified using the Omaha System provide a comprehensive and precise picture of the patient’s situation, therefore, reliable indicators of risk prediction. By mapping the Omaha system through NLP, we could identify risk factors documented in clinical notes, which can save time and effort in extracting information from a large number of notes.

The Environmental domain encompassing social determinants of health was infrequently documented across all clusters, consistent with previous studies.^50,51 In contrast, the Physiological domain was the most frequently documented. While this could be because there were fewer identified risk factors in the Environmental domain compared with the Physiological domain, it is still important to identify social determinants of health, as these factors may be hidden and associated with negative health outcomes.^52,53 The Health-related Behaviors domain such as “Health care supervision” or “Medication regimen” was most frequent in Cluster 2 “Decreased.” Healthcare providers interact with patients during healthcare service and this interaction provides an opportunity for education related to self-management. Home visits over time can have a positive effect on self-care as evidenced by improvements in medication adherence, such as following the recommended dosage and schedule, which falls under the Health-related Behaviors domain. Future work might explore what nursing interventions occur with certain clusters to detect targeted teaching or care management activities.

In Cluster 4 “Steeply Increased,” the Physiological domain was more prevalent than other domain combinations. In other words, physical symptoms were more prominent and noticeable than other risk factors.⁵⁴ Physiological risk factors (eg, abnormal blood pressure, unable to breathe independently, and fever) in patients in HHC can change abruptly over time due to various reasons such as disease progression, medication changes, new medical conditions, or acute illness or injury.^10,55 It is important for HHC providers to monitor patients in Cluster 4 “Steeply Increased,” regularly and be vigilant for any changes in their condition that may require immediate medical attention. In Cluster 2 “Decreased” and Cluster 3 “Decreased and Rebound Increased,” the Environmental domain ranked higher than other domains. This suggests that providers may prioritize having urgent symptoms controlled prior to addressing environmental factors addressed. This would be important to investigate in order to develop targeted interventions that address the underlying causes of the problem and improve health outcomes for individuals within the specific domain by examining the intercorrelations between each domain to gain a comprehensive understanding of how they affect each other. By identifying and addressing these specific risk factors, healthcare providers can better tailor their interventions to the unique needs and circumstances of their patients, ultimately leading to more effective and efficient healthcare delivery.

Lastly, the time difference between the last note used in clustering and the hospital/ED visit varied among patients, with some patients having a hospital/ED visit on the same day as the last note. This suggests that a nurse may identify a patient as high risk or in an emergency during home visits and refer them to 911, leading to an ED visit or hospitalization. On an average, there was a 3-day gap between the last note and the hospital/ED visit. It is essential to note that these time gaps represent the window of opportunity during which a clinical intervention to reduce negative outcomes could be implemented. It is crucial to monitor patients continuously and provide timely interventions to prevent hospitalization.

CONCLUSIONS

This study applied the terms of a standardized nursing language, the Omaha System, using NLP to identify 6 clusters of temporal patterns in documented risk factors that had varying degrees of association with hospitalizations or ED visits. By analyzing for different domains of risk factors, we identified the unique clinical characteristics of each cluster. Future studies can apply our findings to develop cluster-based early warning systems to prevent hospitalizations or ED visits.

Supplementary Material

FUNDING

This work was supported by the Agency for Healthcare Research and Quality (AHRQ) (R01 HS027742), “Building risk models for preventable hospitalizations and ED visits in homecare (Homecare-CONCERN).” The content is solely the responsibility of the authors and does not necessarily represent the official views of the Agency for Healthcare Research and Quality. SHM is supported by the National Institute of Nursing Research (NINR) training grant Reducing Health Disparities through Informatics (RHeaDI) (T32NR007969) as a postdoctoral trainee. MH is supported by the NINR training grant Reducing Health Disparities through Informatics (RHeaDI) (T32NR007969) as a predoctoral trainee and the Jonas Scholarship. SO is supported by the NINR training grant Individualized Care for At Risk Older Adults (T32NR009356) as a postdoctoral trainee.

AUTHOR CONTRIBUTIONS

Study concept and design: JS, SHM, SC, KHB, MVM, and MT. Acquisition of data: MVM, YB, SS, and LE. Analysis and interpretation of data: JS, SHM, KHB, MVM, MH, YB, SO, and MT. Drafting of the manuscript: JS, SHM, SC, KHB, MVM, and MT. Critical revision of the manuscript of important intellectual content: all authors.

SUPPLEMENTARY MATERIAL

Supplementary material is available at Journal of the American Medical Informatics Association online.

CONFLICT OF INTEREST STATEMENT

All authors report no conflicts of interest relevant to this article.

DATA AVAILABILITY

Data available on request due to privacy/ethical restrictions: the data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.