Identification of Patients Admitted With COPD Exacerbations and Predicting Readmission Risk Using Machine Learning

One fifth of patients discharged from hospital for COPD exacerbations are readmitted within 30 days.(1, 3, 4) While therapies and care guidelines have improved, guideline implementation remains poor.(5) Implementing appropriate standards through usual hospital workflow presents significant challenges. One of the top challenges is ensuring all eligible patients with COPD exacerbations are identified in a timely manner.(6) Another top challenge is that staff are often too busy and do not have time to execute evidence-based practices that reduce readmissions.(6) Furthermore, intensive case management can not be offered to everyone because of limited resources. Therefore, it is important that we are able to identify both people who are admitted with COPD early as well as those who are at high risk for readmission.

COPD exacerbations may, at times, not be easily recognized at first and take days to become apparent. Symptoms of exacerbations such as shortness of breath are not specific and signs such as chest radiograph infiltrates can be due to one or more diagnoses. Furthermore, COPD exacerbations can trigger or be triggered by other diseases. As a result, it is not uncommon for admitting physicians to admit patients with multiple provisional diagnoses of heart failure, pneumonia, COPD exacerbation and more. Distinguishing people with COPD exacerbations is further confounded by Electronic Health Records (EHRs) that do not have diagnoses listed as coded elements. The end result is that it is difficult for the rest of the interprofessional team to find COPD patients early in admission. This has been addressed in some U.S. hospitals by having non-health care providers review charts to identify patients admitted with for COPD.(7) An alternate approach has been machine learning and natural language processing. This has been implemented with some success for patients with heart failure but little has been done for people with COPD.(13) In one pilot program, natural language processing helped identify patients admitted with COPD.(7)

To target scarce resources for those who need it most, it would be helpful to further identify patients at high risk of readmission. This would be the first step in determining how to implement effective strategies to reduce readmission rates. There are readmission prediction models developed for medical and surgical patients including the LACE score and the HOSPITAL score.(8, 9) Unfortunately, those that have been studied do not appear to perform well in the COPD population.(10) While factors have been identified that help predict COPD readmission, the models have not been fully validated.(11, 12) The performance could be improved through the use of unstructured data such as clinician progress notes and discharge summaries.

Early identification of people with COPD and knowledge of those who are at risk of readmission can improve health outcomes. Zafar et al. demonstrated that a comprehensive COPD care bundle that consisted of 1. inhaler assessment, 2. appropriate inhaler regimen, 3. early discharge follow up and 4. patient-centered discharge instructions reduced readmissions.(14) Identification of those at high risk of readmission could facilitate enrollment into intensive case management. Therefore, we will conduct the current study to identify patients admitted with acute exacerbations of COPD and stratify patients according to risk of readmission

Methods:

Using retrospective data from the University Health Network (UHN), we will create a data set of admissions to General Internal Medicine for the past 5 years. We estimate this will include approximately 40,000 admissions of which 2,000 will have a most responsible diagnoses of a COPD exacerbation. The data set will contain both structured coded data as well as unstructured text data. Coded data will include age, gender, medications ordered, co-morbidities, laboratory values, and pulmonary function tests. Unstructured text data will include notes in EHR: physician clinic notes, discharge summaries, admission diagnoses, progress notes, and notes from our signover system.

Analysis: We will use several different methods to develop the model including logistic regression, deep neural networks, and convolutional neural networks. Specifically, we will also use statistical machine learning algorithms for event detection using bi-directional long-short term memory neural networks across a variety of input types (e.g., Fourier filter banks, Mel-frequency cepstral coefficients, wavelets, and raw audio). We will also use traditional methods such as dynamic Bayes networks and conditional random fields. On the text analytics side, we will identify key phrases that predict readmission. One approach will be to use discourse analysis to single out "nucleus" phrases from background text. We will also build "joint" predictive models that combine features from the unstructured text and features from the structured coded data. We will use the standard Area under the ROC Curve to assess model performance and use cross validation to minimize the impact of overfitting. Finally, we will then validate our models using a dataset from different centres to determine whether these results are valid and generalizable.

Anticipated results: The development of two validated models based on EHR data: one to accurately identify patients with AECOPD and the second to accurately identify patients at high risk of readmission within 30 days.