Content area
Background: The healthcare industry increasingly values high-quality and personalized care. Patients with heart failure (HF) receiving home health care (HHC) often experience hospitalizations due to worsening symptoms and comorbidities. Therefore, close symptom monitoring and timely intervention based on risk prediction could help HHC clinicians prevent emergency department (ED) visits and hospitalizations. This study aims to (1) describe important variables associated with a higher risk of ED visits and hospitalizations in HF patients receiving HHC; (2) map data requirements of a clinical decision support (CDS) tool to the exchangeable data standard for integrating a CDS tool into the care of patients with HF; (3) outline a pipeline for developing a real-time artificial intelligence (AI)-based CDS tool. Methods: We used patient data from a large HHC organization in the Northeastern US to determine the factors that can predict ED visits and hospitalizations among patients with HF in HHC (9362 patients in 12,223 care episodes). We examined vital signs, HHC visit details (e.g., the purpose of the visit), and clinical note-derived variables. The study identified critical factors that can predict ED visits and hospitalizations and used these findings to suggest a practical CDS tool for nurses. The tool's proposed design includes a system that can analyze data quickly to offer timely advice to healthcare clinicians. Results: Our research showed that the length of time since a patient was admitted to HHC and how recently they have shown symptoms of HF were significant factors predicting an adverse event. Additionally, we found this information from the last few HHC visits before the occurrence of an ED visit or hospitalization were particularly important in the prediction. One hundred percent of clinical demographic profiles from the Outcome and Assessment Information Set variables were mapped to the exchangeable data standard, while natural language processing-driven variables couldn't be mapped due to their nature, as they are generated from unstructured data. The suggested CDS tool alerts nurses about newly emerging or rising risks, helping them make informed decisions. Conclusions: This study discusses the creation of a time-series risk prediction model and its potential CDS applications within HHC, aiming to enhance patient outcomes, streamline resource utilization, and improve the quality of care for individuals with HF. Clinical Relevance: This study provides a detailed plan for a CDS tool that uses the latest AI technology designed to aid nurses in their day-to-day HHC service. Our proposed CDS tool includes an alert system that serves as a guard rail to prevent ED visits and hospitalizations. This tool can potentially improve how nurses make decisions and improve patient outcomes by providing early warnings about ED visits and hospitalizations.
Abstract
Background: The healthcare industry increasingly values high-quality and personalized care. Patients with heart failure (HF) receiving home health care (HHC) often experience hospitalizations due to worsening symptoms and comorbidities. Therefore, close symptom monitoring and timely intervention based on risk prediction could help HHC clinicians prevent emergency department (ED) visits and hospitalizations. This study aims to (1) describe important variables associated with a higher risk of ED visits and hospitalizations in HF patients receiving HHC; (2) map data requirements of a clinical decision support (CDS) tool to the exchangeable data standard for integrating a CDS tool into the care of patients with HF; (3) outline a pipeline for developing a real-time artificial intelligence (AI)-based CDS tool.
Methods: We used patient data from a large HHC organization in the Northeastern US to determine the factors that can predict ED visits and hospitalizations among patients with HF in HHC (9362 patients in 12,223 care episodes). We examined vital signs, HHC visit details (e.g., the purpose of the visit), and clinical note-derived variables. The study identified critical factors that can predict ED visits and hospitalizations and used these findings to suggest a practical CDS tool for nurses. The tool's proposed design includes a system that can analyze data quickly to offer timely advice to healthcare clinicians.
Results: Our research showed that the length of time since a patient was admitted to HHC and how recently they have shown symptoms of HF were significant factors predicting an adverse event. Additionally, we found this information from the last few HHC visits before the occurrence of an ED visit or hospitalization were particularly important in the prediction. One hundred percent of clinical demographic profiles from the Outcome and Assessment Information Set variables were mapped to the exchangeable data standard, while natural language processing-driven variables couldn't be mapped due to their nature, as they are generated from unstructured data. The suggested CDS tool alerts nurses about newly emerging or rising risks, helping them make informed decisions.
Conclusions: This study discusses the creation of a time-series risk prediction model and its potential CDS applications within HHC, aiming to enhance patient outcomes, streamline resource utilization, and improve the quality of care for individuals with HF.
Clinical Relevance: This study provides a detailed plan for a CDS tool that uses the latest Al technology designed to aid nurses in their day-to-day HHC service. Our proposed CDS tool includes an alert system that serves as a guard rail to prevent ED visits and hospitalizations. This tool can potentially improve how nurses make decisions and improve patient outcomes by providing early warnings about ED visits and hospitalizations.
KEYWORDS
artificial intelligence, clinical decision support, electronic health records, heart failure, home care services, nursing informatics
INTRODUCTION
Heart failure (HF) is a prevalent, expensive, and debilitating condition that affects over 6.2 million adults in the United States, leading to annual expenses exceeding $30 billion (Young, 2020). HF often results in patients requiring emergency department (ED) visits or hospitalizations when the condition is not managed effectively (Collins et al., 2013; Farré et al., 2017). Despite efforts to improve the transition from hospital to home, the 30-day hospitalization rate after being discharged for any reason for HF remains high at approximately 23% (Madanat et al., 2021). Home health care (HHC) provides post-acute care for older adults in their homes, including for those with HF. Preventing avoidable hospitalizations and ED visits is critical for reducing the burden on patients and their caregivers and lowering healthcare costs. Several previous studies have tested clinician, clinic, and health system interventions aimed at enhancing care quality for HHC HF patients (Jones et al., 2017), including the development of clinical decision support (CDS) to assist clinicians in making post-acute HHC referrals, resulting in significant reductions in 30- and 60-day readmissions (Bowles etal, 2015).
Research suggests that CDS tools integrated within the electronic health record (EHR) can analyze patient data and trigger timely, actionable alerts prompting early interventions to prevent unplanned hospitalization for patients with HF (Lai et al., 2020; Romero-Brufau et al., 2020). HHC clinicians assess patients' status and vital signs to create care plans and document them in the HHC EHR. According to a recent scoping review, more than 20 studies have used machine learning (ML) methods to predict adverse outcomes, such as hospitalization and mortality, in HHC settings (Hobensack et al., 2023). However, although many predictive models had been developed for nurses to use, few have been implemented in real-world HHC settings. This is partly because of a lack of clear guidance on how those models should be integrated into routine nursing practices (Wieben et al, 2023).
Our team created a time-series risk model to forecast ED visits and hospitalizations in HF patients by analyzing longitudinal HHC data (Chae et al., 2023). We found that the best-performing model was developed by integrating variables from the Outcome and Assessment Information Set (OASIS)-a thorough evaluation performed at the beginning of an HHC episode-together with data from vital signs, visit details, and clinical notes collected within 4 days prior to an ED visit or hospitalization. In the current paper, we extend the application of our preliminary predictive model, to the investigation of potential CDS application in HHC that could improve patient outcomes, optimize resource use, and enhance quality of care for patients with HF.
Until recently a lack of data interoperability, which refers to the capacity for different information technology systems and software applications to communicate, exchange, interpret, and use data seamlessly, presented significant challenges for implementation of CDS systems (Kennedy et al., 2021). However, developments in healthcare data interoperability standards provide a promising avenue for successful implementation into the existing EHR. The Fast Healthcare Interoperability Resources (FHIR) standard is gaining widespread adoption and is currently mandated as a significant standard for healthcare information exchange globally (Ayaz et al., 2021). This paper outlines how established standards and technologies can be used to implement a CDS solution that aligns the goals of payers, clinicians, and patients to improve patient outcomes (Mandel et al., 2016). This work could further help streamline the implementation of Al-driven risk models to improve patient outcomes.
This study bridges the gaps in the literature regarding development of CDS tools in HHC patients and focuses on three primary objectives: to (1) describe critical variables associated with a higher risk of ED visits and hospitalizations among HHC patients with HF; (2) map data requirements of a CDS tool to the exchangeable data standard for integrating a CDS tool into the care of patients with HF; and (3) present the outline of a pipeline applicable to establish a realtime, AI-based, CDS tool.
DESIGN
Materials and methods
HHC setting and study period
This research utilizes data from a large, non-profit HHC organization in the Northeastern US known for its extensive service to patients at home for over 130 years. The organization's frontline team, composed of nurses, social workers, and rehabilitation therapists, offers a broad array of HHC services for post-hospital care or ongoing home-based health maintenance.
Our team has created a time series model for future risk detection and now aim to integrate the model as a CDS tool in the EHR system (Chae et al., 2023). The risk model was developed for postacute adults (aged >18 years) with a HF diagnosis [the International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) codes: 150.x, 111.0, 113.0, 113.1, 113.2] (World Health Organization, 2019). Subjects were admitted to the HHC agency between January 1, 2015 and December 31, 2017. For prediction, the model uses seven sets of features (Chae et al., 2023). Three features were from structured data, which is organized and stored in a format (OASIS data items, which include clinical demographic profiles and patient needs, vital signs, and visit details). The other four features were drawn from the unstructured data, which is not organized ina predefined format and is often presented in free-text or narrative form (i.e., clinical documentation generated by HHC professionals). This data was processed using natural language processing (NLP), numerical representations based on occurrence frequency, deep learning, and topic modeling.
Identifying critical variables associated with a higher risk of ED visits and hospitalizations among patients with HF in HHC
As a post-hoc analysis, we employed Shapley additive explanations (SHAP) (Lundberg & Lee, 2017) to interpret how different variables influence the accuracy of the model's ED visits and hospitalization risk predictions. In essence, SHAP uses ML models to identify the importance of additive variables (Lundberg & Lee, 2017). We implemented a decision tree-based (LightGBM version) Python SHAP package (https://github.com/slundberg/shap) to determine the contribution of each variable to the risk of ED visits or hospitalizations. The top 20 variables were identified using the mean absolute SHAP values.
Data requirements for the CDS
To run the CDS model for decision support, the variables must be retrieved from the patient record. We linked the model variables to FHIR, highlighting how the information should be structured according to the FHIR standards, a standard for healthcare data exchange published by HL7® (Health Level 7 FHIR, 2024). We detail the relevant FHIR attributes, the constraints on the data we utilize, and the acceptable value sets that enable the CDS tool to function seamlessly within the EHR system. This approach ensures the tool is equipped with the necessary information to support HHC clinicians to identify patients with higher clinical concern and to provide timely interventions.
A pragmatic guide to establishing CDS governance
For the proposed CDS tool, we adhered to the core principles of "five rights": the right information to the right person in the right intervention format through the right channel at the right time in the clinical workflow (Campbell, 2013). Specifically, we ensured the delivery of appropriate content (such as OASIS, vital signs, and clinical notes) to the correct stakeholders (patients and clinicians) at the most opportune times (during actual visits or before specific events), through the right channels (such as the EHR's messaging portal), and in an accessible format (message data presentation) (Osheroff et al, 2012). The system architecture illustrated how the CDS logic will trigger in three stages.
RESULTS
Overall, we identified a total of 9362 patients with HF who received 176,209 visits across 12,223 episodes о? HHC. Most patients were women (61%), with an average age of 81.7 years [standard deviation (SD) of 11 years] at the beginning of their care. Their average length of stay in HHC was 48 days (SD of 56 days). Approximately one out of every four patients required hospitalization or visited the ED within 60 days of HHC admission.
Variables associated with risk for ED visits and hospitalization
We identified features that significantly contributed to the classification task based on their SHAP values. SHAP values were assigned numbers of importance for each feature of the model. The positive or negative sign of a SHAP value of a feature determined whether the feature has positive or negative impact on the outcome, and its magnitude showed the degree of the impact on the outcome. Figure 1 represents the top 20 features, ordered in the magnitude of the features' SHAP values, associated with the ED visit or hospitalization risk as the outcome within the 4 days prior to its occurrence.
The number of days since HHC admission to the current visit was the feature most significantly associated with risk of ED visits or rehospitalization (Figure 1). The risk is higher when this period is more brief (negative SHAP value). Similarly, the second, fifth, and eighth most significant features-that is, the numbers of days between the last visit and two visits previously, three visits previously, and four visits previously, respectively-reveal that shorter gaps between the adjacent visits correspond to higher incidence of ED or rehospitalization (negative SHAP value). The third and tenth features highlight a clear link: the more recent and frequent the documented evidence of HF symptoms, the higher the clinical concern for ED visits and hospitalizations. The third-ranked feature, the total number of HF symptoms noted during the last visit, was extracted through NLP. Similarly, the tenth-ranked feature accounts for the "total number of HF symptoms identified five visits prior to the event occurring." This pattern suggests that the risk increases with the recency and frequency of documented HF symptoms.
The fourth-ranked feature focuses on the "visit purpose," encompassing various reasons such as psychiatric care, missed visits (e.g., no visit, patient was absent, or patient refused the visit), complex care and infusions, rehabilitation, nursing training and education, and general nursing assessments and care. This feature suggests that the nature of a current or previous HHC visit indicates the likelihood of an ED visit or hospitalization. Specifically, our detailed analysis revealed that patients who had missed visits-instances where HHC was scheduled but did not occur because the patient was not at home, potentially because they were in the hospital, or the door remained unanswered-were at a greater clinical concern of negative outcomes compared to those with no missed visit.
Among the other significant predictors, we identified six features derived from NLP using the deep learning model (Erickson et al., 2020), a topic labeled "Instruction on self-management" from our topic modeling, and three terms identified by their frequency through term frequency-inverse document frequency (TF-IDF) » analysis. The three terms were "ER (emergency room)" "no further," and "generic." TF-IDF is a way to measure the importance of words in a document compared to their importance in a larger collection of documents, like a set of articles, books, or web pages; it helps find words that are significant within a specific document (Lubis et al., 2021).
The "Instruction on self-management" topic comprises words ranked by the frequency of their occurrence, including terms related to patient care and instructions provided by healthcare professionals, such as "patient (pt)," "visiting nurse service (VNS)," and directives regarding weight management and medication adherence. Traditional clinical indicators like categorical vital signs [e.g., for blood pressure: O = missing, 1- normal (less than 120/80 mm Hg), 2=elevated (systolic between 120 and 129 mm Hg and diastolic less than 80mm Hg), 3=stage 1 hypertension (systolic between 130 and 139 mm Hg or diastolic between 80 and 89 mm Hg), 4=stage 2 hypertension (systolic at least 140mm Hg or diastolic at least 90mm Hg), or stage 5=hypertensive crisis (systolic over 180 mm Hg and/or diastolic over 120mm Hg)] (Greenland & Peterson, 2017) were not among the top 20 predictors.
Our analysis found that the optimal prediction window for hospitalizations and ED visits was within 4 days, corroborating the findings from our previous study (Chae et al., 2023). Overall, nine out of the top 20 features that were most indicative of clinical concern, that is, the first, second, third, fourth, fifth, sixth, eighth, tenth, and thirteenth-ranked features, were derived from the most recent four HHC visits, providing a focused snapshot of the predictive data (as depicted in Figure 1).
Data requirements
Our CDS application, designed for HF patients, utilizes FHIR tools to extract relevant data elements from the EHR. These elements are detailed in Table 2, with value sets in Table 3. Additionally, 76 data elements crucial for the risk model, mapped to the FHIR data elements, are in Table 4. As a result, 100% of clinical demographic profiles from OASIS variables were mapped (91% were fully mapped and 9% were partially mapped). Visit characteristics were partially mapped. However, NLP-driven variables could not be mapped due to their nature, as they are generated from unstructured data, that is, clinical notes using NLP. In a previous study, the addition of rule-based, NLP-derived variables boosted the predictive model's performance (F1 score) from 0.57 to 0.67, representing a 10% enhancement compared to the baseline (Chae et al., 2023).
High-level architecture and real-time implementation
We propose a high-level architecture for a time series risk model that creates a step-by-step process for HHC clinicians to receive alerts and CDS. A summary of the five rights for the targeted CDS component is presented in Table 1.
The technical architecture supporting the real-time NLP step-bystep process CDS tool would integrate cutting-edge and emerging technological features. Figure 2 provides an overview of the NLPdriven CDS system's infrastructure, which involves extracting notes from the EHR, organizing them, and inputting them into an NLP step-by-step process to extract information, and finally, inputting the processed text features into the time-series risk models. The resulting risk scores from the time-series risk models are delivered to clinicians as alert messages.
In Step 1 (HHC Visit Stage), intake staff collect preliminary information from the referral source (e.g., reason for referral, some clinical information, insurance information/confirmation of home care eligibility). Once a case is accepted and it goes to the team for scheduling and assignment, the HHC clinician (mostly a nurse but sometimes a physical therapist) then completes the first visit (OASIS), and after this is done then the case is considered admitted. The EHR CDS tool trigger is processed after the first visit when both OASIS and visit notes are available for the model. In this stage, the HHC scheduler will accept and confirm the appointment by making an entry in the EHR system. The appointment entry triggers the EHR CDS tool to check the patient record for a predefined list of Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT) or ICD-10 codes indicating HF-related conditions.
In Step 2 (the CDS Activation Stage), if the patient record has a match for HF conditions, a message containing a link to the relevant information is sent to the HHC clinician's EHR CDS application. The clinician reviews the CDS application message and the information. If the clinician declines the enrollment option, the CDS tool workflow will not be activated. Once the clinician accepts enrollment, the EHR CDS tool requests the patient's previous clinical notes and OASIS from the EHR through the FHIR server (that has been updated and optimized for use of unstructured data elements). The EHR CDS tool employs a Substitutable Medical Applications and Reusable Technologies (SMART) on the FHIR web service to retrieve both structured data elements (such as OASIS items, vital signs, and visit characteristics) and unstructured data elements (including visit notes and care coordination notes) for the target patient from the EHR system. Within a cloud computing environment, an NLP engine processes clinical notes from the EHR, and structured data elements are extracted from the OASIS assessment conducted after the first visit of the HHC episode and from the vital signs and visit patterns that are recorded during the routine provision of HHC services. An Al-powered interface engine (AutoGluon) receives various input variables and stores the results in a web-based platform (Erickson et al., 2020). A custom computer program within this platform performs text extraction and linguistic feature engineering at regular intervals, associating concept unique identifiers with patient identifiers. These identifiers are used as input to the Al-powered interface engine (AutoGluon) at the HHC visit level. Prediction probabilities and classifications are stored in a table within the web-based platform. A message call from the EHR cloud determines whether the ML model's cutoff threshold was reached or not; if reached over the threshold, an alert will be triggered.
In Step 3 (the Care Management Stage), the EHR cloud sends a request to a web-based platform to generate a risk score. This risk score is then sent back to the EHR cloud and presented to the clinician as an alert when they access the patient's chart in the EHR system. The HHC clinician reviews the CDS message, risk score, and other relevant information before visiting the patient. Upon receiving the alert, the clinician examines the patient's condition and current HF treatments and medications. Supported by the current findings, the clinician will determine the need to perform additional tests or start early intervention protocols. During the episode, the HHC clinician may discuss with the patient their current treatments and medications for HF and how to refine the selection of drugs, formulations, and dosages for optimal outcomes.
DISCUSSION
This study's major and novel contribution is identifying risk factors associated with time series risk for ED visits and hospitalizations within the HHC setting and suggesting a workflow for CDS implementation. Applying the variable selection technique (SHAP values), we found a high correlation between the frequency of recent HHC visits and the patient's risk identification for ED visits and hospitalization. Specifically, we found that more frequent HHC visits and shorter lengths of HHC episodes are closely associated with the risk identification of ED visits and hospitalizations. The more frequent visits over a short period of time might have been due to a greater need for care and attention. This is consistent with findings that frequent clinician visits often correlate with higher clinical complexity or worsening patient health status (Rossetti et al., 2021). This is also observed in the hospital setting, where shorter interval between clinician assessments can signal early deterioration (Keim-Malpass et al., 2020; Schnock et al., 2021). Previous research aligns with the current findings, indicating that patients with shorter HHC episodes (defined as fewer than 21 days), when compared to a length of stay of 21 days or more, were more likely to be readmitted (O'Connor et al., 2015). Our primary outcome of interest was hospitalization or ED visit anytime within 60 days of HHC admission, therefore, the shorter lengths of stay in HHC in this study could be indicative of future hospitalization. Additionally, we found that missed HHC visits are linked to increased clinical concern, aligning with previous studies that show missing, delaying, or refusing HHC services is associated with increased patient risk (Topaz et al., 2015). We did not adjust for missed visits that were due to a patient's ED visit or hospitalization because we do not have access to that information. This finding indicates the needs for educating the patients and clinicians that skipping visits increases patients' risk of ED visits or hospitalizations. HHC care patterns and visit characteristics should be strongly considered in developing further risk prediction modeling.
We also observed a strong correlation between several rulebased NLP-derived features and clinical concern. The first set of NLP features includes the total number of specific HF-related symptoms extracted from clinical notes during the current and previous HHC visits. Previous research indicates that this factor is crucial for predicting risk at the HHC episode level (Chae et al., 2022; Song, Hobensack, et al., 2022; Song, Ojo, et al., 2022), and this study reinforces its significance in time series risk modeling. In addition, we found that the presence of a topic model indicating the patient's instruction on self-management was potentially due to or results in higher risk for ED visits or hospitalizations. These findings align with previous studies that show that higher symptom burden and lower self-management skills among patients with HF are indicators of risk for negative outcomes (Scháfer-Keller et al., 2021). Our study emphasizes the critical need for timely interventions, including thorough assessments, patient education, and management of HF symptoms within specific time frames during HHC episodes to reduce the risk of future ED visits or hospitalizations.
Additionally, we discovered that several deep learning and TFIDF-derived features were not readily interpretable as clinical risk factors; six of the 20 highly correlated risk features that are deep learning-based would not be easily understood by HHC clinicians. This is because deep learning models are often seen as "black box" models due to their complexity, making it difficult to provide simple and intuitive explanations. Furthermore, TF-IDF analysis identified high-risk words such as "ER" (indicating previous ED visits) and lowrisk words like "no further" (suggesting no additional HHC is needed) as relevant features for predicting ED visits and hospitalizations. Developing a clinically interpretable risk prediction model is crucial for clinical adoption and implementation. It builds trust in decisionmaking, facilitates error detection and correction, and supports integration into clinical workflows (Markus et al., 2021). To inspire user trust in the model, AI development should communicate the performance with transparency, present confidence, report uncertainties appropriately, explain what is behind the model outputs, acknowledge limitations, and consider human error in the loop (Vôssing et al., 2022). In principle, the risk can be partially mitigated through human-centered design, with human users at the forefront of the design and operation of AI systems. Further research is needed to determine the best way to present these risk factors to HHC clinicians.
In hospital settings, trends in vital signs often provide strong indicators for risk prediction, with many hospital-based risk models relying heavily on these routinely collected measurements (Gerry et al., 2020). Surprisingly, our study did not find vital signs among the top 20 features highly correlated with HHC patient risk. This suggests that vital signs in HHC allow only limited view of changes in the patient's condition because they are usually collected only every few days compared to practices in acute hospitals, where vital signs are monitored continuously or hourly. For patients with HF, changes in vital signs might be more significant if they deviate sharply from abnormal baseline values. Moreover, sudden or gradual changes in vital signs are often among the last indicators of significant deterioration that occurs in patients with HF that need intervention (Cardona-Morrell et al., 2015), suggesting that the visit-level time window might be too broad to detect these critical changes effectively.
We also found nine important features for predicting outcomes were generated within four visits before the event (Figure 1). This highlights the fact that clinicians need to focus on patients recently admitted to HHC regarding the risk of ED visits and hospitalizations. We may require additional clinical measurements from HHC visits to detect meaningful changes in a patient's health status. Additionally, other data resources, such as telehealth data from patients with HF, would be beneficial, especially from HHC agencies that have telehealth programs.
To support a seamless integration of the CDS with the system in use and an effective implementation that does not disrupt the current HHC workflows, the timing and availability of the appropriate tools is critical. Our proposed tool required several data elements retrieved from the EHR using special tools called FHIR, which are like keys that help us access the data we need. We illustrated the data requirement specifications, FHIR, relevant FHIR attributes, and the value set for the CDS tool.
We mapped variables in the risk model to applicable vocabulary standards using FHIR. One hundred percent of OASIS variables were partially or fully mapped, and visit characteristics were all partially mapped, while NLP-driven variables could not be mapped. We recommend mapping NLP-driven variables to standardized terminologies like the Omaha System, Logical Observation Identifiers Names and Codes (LOINC), or SNOMED-CT.
One of our challenges was the lack of data elements in FHIR specific to HHC. Of note, during data collection at the HHC agency for this study, we noted that, for the same data element, different health systems might use different coding terminologies or do not have any available terminologies. The consequences of not being able to map all the variables from a HHC setting to the FHIR standard can indeed impact the functionality and utility of CDS tool. To mitigate these consequences, it's essential for HHC agencies to work collaboratively with EHR vendors, standards organizations, and other stakeholders to enhance data interoperability and ensure that important data elements are included in the FHIR standard or through custom extensions. This will support future implementation of the CDS tool and data-driven decision-making in the HHC setting.
Another challenge or limitation in real-world implementation within HHC environments is that while incorporating ML into risk prediction systems enhances the accuracy of risk assessments, it also introduces difficulties related to interpretability and gaining the trust of clinicians (Nundy et al., 2019).
Our proposed CDS tool includes an alert system that serves as a guard rail to prevent ED visits and hospitalizations. To ensure the alert system is effective and usable, it should be designed to minimize unnecessary disruption to clinical workflows, provide clear instructions, include relevant information for the clinicians, and be tested and refined with clinicians as real-world users (Olakotan & Yusof, 2020). Our next steps involve proposing a simulation using a cohort of HF patients in HHC to examine how frequently the CDS alert will be triggered in the EHR CDS.
CONCLUSIONS
ED use and readmission rates for patients with HF receiving HHC services remain stubbornly high (Sterling Madeline et al., 2020), highlighting the essential role of the proposed CDS tool in practice. The time series risk model we developed indicates that HHC clinicians can pinpoint patients at risk of ED visits or hospitalization within a four-day window before such events, enabling the provision of timely, preventive interventions. Furthermore, our prior research underscores the significance of rule-based, NLPderived features in predicting risk at the level of HHC episodes (Jiyoun Song, М. Ojo, et al., 2022), and this study reaffirms their value in temporal risk modeling. Therefore, the prediction model, particularly when considering NLP-derived features, can serve as a catalyst for initiating telemedicine consultations, paramedic programs, medication adjustments, or face-to-face outpatient visits with healthcare clinicians. This study discusses the creation of a time-series risk prediction model and its potential CDS applications within HHC, aiming to enhance patient outcomes, streamline resource utilization, and improve the quality of care for individuals with HF.
REFERENCES
Ayaz, M., Pasha, M. F., Alzahrani, M. Y., Budiarto, R., & Stiawan, D. (2021). The fast health interoperability resources (FHIR) standard: Systematic literature review of implementations, applications, challenges and opportunities [review]. JMIR Medical Informatics, 9(7), е21929. https://doi.org/10.2196/21929
Bowles, K. H., Chittams, J., Heil, E., Topaz, M., Rickard, K., Bhasker, M., Tanzer, M., Behta, M., & Hanlon, А. L. (2015). Successful electronic implementation of discharge referral decision support has a positive impact on 30- and 60-day readmissions. Research in Nursing & Health, 38(2), 102-114. https://doi.org/10.1002/nur.21643
Campbell, R. (2013). The five" rights" of clinical decision support. Journal of AHIMA, 84(10), 42.
Cardona-Morrell, M., Nicholson, M., & Hillman, К. (2015). Vital signs: From monitoring to prevention of deterioration in general wards. In Vincent JL. (eds) Annual Update in Intensive Care and Emergency Medicine. Springer, Cham. https://doi.org/10.1007/978-3-319-13761-2_39
Chae, S., Davoudi, A., Song, J., Evans, L., Hobensack, M., Bowles, K. H., McDonald, M. V., Barrón, Y., Rossetti, S. C., Cato, K., Sridharan, S., & Topaz, M. (2023). Predicting emergency department visits and hospitalizations for patients with heart failure in home healthcare using a time series risk model. Journal of the American Medical Informatics Association, 129, 1622-1633. https://doi.org/10.1093/jamia/ocad129
Chae, S., Song, J., Ojo, M., Bowles, K. H., McDonald, M. V., Barrón, Y., Hobensack, M., Kennedy, E., Sridharan, S., Evans, L., & Topaz, M. (2022). Factors associated with poor self-management documented in home health care narrative notes for patients with heart failure. Heart & Lung, 55, 148-154. https://doi.org/10.1016/j.hrting.2022. 05.004
Collins, S. P., Pang, P. S., Fonarow, G. C., Yancy, C. W., Bonow, R. O., & Gheorghiade, M. (2013). Is hospital admission for heart failure really necessary? The role of the emergency department and observation unit in preventing hospitalization and rehospitalization. Journal of the American College of Cardiology, 61(2), 121-126.
Erickson, N., Mueller, J., Shirkov, A., Zhang, H., Larroy, P., Li, M., & Smola, A. (2020). AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data.
Farré, N., Vela, E., Cleries, M., Bustins, M., Cainzos-Achirica, M., Enjuanes, C., Moliner, P., Ruiz, S., Verdú-Rotellar, J. M., & ComínColet, J. (2017). Real world heart failure epidemiology and outcome: A population-based analysis of 88,195 patients. PLoS One, 12(2), e0172745. https://doi.org/10.1371/journal.pone.0172745
Gerry, S., Bonnici, T., Birks, J., Kirtley, S., Virdee, P. S., Watkinson, P. J., & Collins, G. S. (2020). Early warning scores for detecting deterioration in adult hospital patients: Systematic review and critical appraisal of methodology. BMJ, 369, m1501. https://doi.org/10.1136/bmj.m1501
Greenland, P., & Peterson, E. (2017). The new 2017 ACC/AHA guidelines "up the pressure" on diagnosis and treatment of hypertension. JAMA, 318(21), 2083-2084. https://doi.org/10.1001/jama.2017.18605
Health Level 7 FHIR. (2024). Retrieved January 5. https://www.hl7.org/ fhir/
Hobensack, M., Song, J., Scharp, D., Bowles, K. H., & Topaz, M. (2023). Machine learning applied to electronic health record data in home healthcare: A scoping review. International Journal of Medical Informatics, 170, 104978. https://doi.org/10.1016/j.iimedinf.2022.104978
Jones, C. D., Bowles, K. H., Richard, A., Boxer, R. S., & Masoudi, F. A. (2017). High-value home health care for patients with heart failure. Circulation. Cardiovascular Quality and Outcomes, 10(5), e003676. https://doi.org/10.1161/CIRCOUTCOMES.117.003676
Keim-Malpass, J., Clark, M. T., Lake, D. E., & Moorman, J. К. (2020). Towards development of alert thresholds for clinical deterioration using continuous predictive analytics monitoring. Journal of Clinical Monitoring and Computing, 34(4), 797-804. https://doi.org/10. 1007/s10877-019-00361-5
Kennedy, E. E., Bowles, K. H., & Aryal, S. (2021). Systematic review of prediction models for postacute care destination decision-making. Journal of the American Medical Informatics Association, 29(1), 176186. https://doi.org/10.1093/jamia/ocab197
Lai, H.-J., Tan, T.-H., Lin, C.-S., Chen, Y.-F., & Lin, H.-H. (2020). Designing a clinical decision support system to predict readmissions for patients admitted with all-cause conditions. Journal of Ambient Intelligence and Humanized Computing, 1-10. https://doi.org/10.1007/s12652-01901579-6
Lubis, A. R., Nasution, M. K., Sitompul, O. S., & Zamzami, E. M. (2021). The effect of the TF-IDF algorithm in times series in forecasting word on social media. Indones. J. Electr. Eng. Comput. Sci, 22(2), 976.
Lundberg, S. M., & Lee, S.-1. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30, 4765-4774.
Madanat, L., Saleh, M., Maraskine, M., Halalau, A., & Bukovec, F. (2021). Congestive heart failure 30-day readmission: Descriptive study of demographics, Co-morbidities, heart failure knowledge, and selfcare. Cureus, 13(10), e18661. https://doi.org/10.7759/cureus.18661
Mandel, J. C., Kreda, D. A., Mandl, К. D., Kohane, |. S., & Ramoni, В. В. (2016). SMART on FHIR: A standards-based, interoperable apps platform for electronic health records. Journal of the American Medical Informatics Association, 23(5), 899-908.
Markus, A. F., Kors, J. A., & Rijnbeek, P. R. (2021). The role of explainability in creating trustworthy artificial intelligence for health care: A comprehensive survey of the terminology, design choices, and evaluation strategies. Journal of Biomedical Informatics, 113, 103655. https://doi.org/10.1016/j.jbi.2020.103655
Nundy, S., Montgomery, T., & Wachter, R. M. (2019). Promoting trust between patients and physicians in the era of artificial intelligence. JAMA, 322(6), 497-498. https://doi.org/10.1001/jama.2018.20563
O'Connor, M., Hanlon, A., Naylor, M. D., & Bowles, К. H. (2015). The impact of home health length of stay and number of skilled nursing visits on hospitalization among medicare-reimbursed skilled home health beneficiaries. Research in Nursing & Health, 38(4), 257-267. https://doi.org/10.1002/nur.21665
Olakotan, O. O., & Yusof, M. M. (2020). Evaluating the alert appropriateness of clinical decision support systems in supporting clinical workflow. Journal of Biomedical Informatics, 106, 103453. https:// doi.org/10.1016/j.jbi.2020.103453
Osheroff, J. A., Teich, J., Levick, D., Saldana, L., Velasco, F., Sittig, D., Rogers, K., & Jenders, R. (2012). Mproving outcomes with clinical decision support: An Implementer's guide (2nd ed.). Himss Publishing. https://doi.org/10.4324/9780367806125
Romero-Brufau, S., Wyatt, K. D., Boyum, P., Mickelson, M., Moore, M., & Cognetta-Rieke, C. (2020). Implementation of artificial intelligencebased clinical decision support to reduce hospital readmissions at a regional hospital. Applied Clinical Informatics, 11(4), 570-577. https://doi.org/10.1055/s-0040-1715827
Rossetti, S. C., Knaplund, C., Albers, D., Dykes, P. C., Kang, М. J., Korach, T. Z., Zhou, L., Schnock, K., Garcia, J., Schwartz, J., Fu, L. H., Klann, J. G., Lowenthal, G., & Cato, K. (2021). Healthcare process modeling to phenotype clinician behaviors for exploiting the signal gain of clinical expertise (HPM-ExpertSignals): Development and evaluation of a conceptual framework. Journal of the American Medical Informatics Association, 28(6), 1242-1251. https://doi.org/10.1093/jamia/ocab006
Scháfer-Keller, P., Santos, С. C., Denhaerynck, K., Graf, D., Vasserot, K., Richards, D. A., & Stromberg, А. (2021). Self-care, symptom experience, needs, and past health-care utilization in individuals with heart failure: Results of a cross-sectional study. European Journal of Cardiovascular Nursing, 20(5), 464-474. https://doi.org/10.1093/eurjcn/zvaa026
Schnock, К. O., Kang, М. J., Rossetti, 5. C., Garcia, J., Lowenthal, G., Knaplund, C., Chang, F., Albers, D., Korach, T. Z., Zhou, L., Klann, J. G., Cato, K., Bates, D. W., & Dykes, P. С. (2021). Identifying nursing documentation patterns associated with patient deterioration and recovery from deterioration in critical and acute care settings. International Journal of Medical Informatics, 153, 104525. https:// doi.org/10.1016/j.iimedinf.2021.104525
Song, J., Hobensack, M., Bowles, K. H., McDonald, M. V., Cato, K., Collins Rossetti, S., Chae, S., Kennedy, E., Barrón, Y., Sridharan, S., & Topaz, M. (2022). Clinical notes: An untapped opportunity for improving risk prediction for hospitalization and emergency department visit during home health care. Journal of Biomedical Informatics, 128, 104039. https://doi.org/10.1016/j.jbi.2022.104039
Song, J., Ojo, M., Bowles, К. H., McDonald, M. V., Cato, K., Rossetti, S. C., Adams, V., Chae, S., Hobensack, M., Kennedy, E., Tark, A., Kang, M. J., Woo, K., Barrón, Y., Sridharan, S., & Topaz, М. (2022). Detecting language associated with home health care Patient's risk for hospitalization and emergency department visit. Nursing Research, 71, 285-294. https://doi.org/10.1097/nnr.0000000000 000586
Sterling Madeline, R., Kern Lisa, M., Safford Monika, M., Jones Christine, D., Feldman Penny, H., Fonarow Gregg, C., Sheng, S., Matsouaka Roland, A., DeVore Adam, D., Lytle, B., Xu, H., Allen Larry, A., Deswal, A., Yancy Clyde, W., S Albert Nancy, M. (2020). Home health care use and post-discharge outcomes after heart failure hospitalizations. JACC: Heart Failure, 8(12), 1038-1049. https://doi.org/10.1016/j.jchf. 2020.06.009
Topaz, M., Kang, Y., Holland, D. E., Ohta, B., Rickard, K., & Bowles, К. H. (2015). Higher 30-day and 60-day readmissions among patients who refuse post acute care services. The American Journal of Managed Care, 21(6), 424-433.
Vossing, M., Kühl, N., Lind, M., & Satzger, С. (2022). Designing transparency for effective human-Al collaboration. Information Systems Frontiers, 24(3), 877-895. https://doi.org/10.1007/s10796-022-10284-3
Wieben, A. M., Walden, R. L., Alreshidi, В. G., Brown, 5. F., Cato, K., Coviak, С. P., Cruz, C., D'Agostino, F., Douthit, В. J., Forbes, T. H., 3rd, Gao, G., Johnson, S. G., Lee, М. A., Mullen-Fortino, M., Park, J. 1., Park, S., Pruinelli, L., Reger, A., Role, J., & Jeffery, A. D. (2023). Data science implementation trends in nursing practice: A review of the 2021 literature. Applied Clinical Informatics, 14(3), 585-593. https://doi.org/10.1055/a-2088-2893
World Health Organization (Ed.). (2019). International statistical classification of diseases and related health problems (11th ed.). World Health Organization.
Young, B. A. (2020). Health disparities in advanced heart failure treatment: The intersection of race and sex. JAMA Network Open, 3(7), е2011034. https://doi.org/10.1001/jamanetworkopen.2020.11034
Copyright Blackwell Publishing Ltd. Jan 2025