Elsevier

Computers in Biology and Medicine

Volume 101, 1 October 2018, Pages 199-209
Computers in Biology and Medicine

Predicting hospital readmission for lupus patients: An RNN-LSTM-based deep-learning methodology

https://doi.org/10.1016/j.compbiomed.2018.08.029Get rights and content

Highlights

  • Lupus is a chronic autoimmune disease that impacts several system organ classes.

  • Predicting hospital readmissions for Lupus patients is a challenging task.

  • Longitudinal EHR and Deep Learning models are used to predict readmission of Lupus patients.

  • Sequential deep learning methods help capture the temporal relationships in the data.

  • Deep learning method of Recurrent Neural Networks – Long Short Term Memory produces the best prediction results with an AUC of 0.70 on holdout sample.

Abstract

Hospital readmission is one of the critical metrics used for measuring the performance of hospitals. The HITECH Act imposes penalties when patients are readmitted to hospitals if they are diagnosed with one of the six conditions mentioned in the Act. However, patients diagnosed with lupus are the sixth highest in terms of rehospitalization. The heterogeneity in the disease and patient characteristics makes it very hard to predict rehospitalization. This research utilizes deep learning methods to predict rehospitalization within 30 days by extracting the temporal relationships in the longitudinal EHR clinical data. Prediction results from deep learning methods such as LSTM are evaluated and compared with traditional classification methods such as penalized logistic regression and artificial neural networks. The simple recurrent neural network method and its variant, gated recurrent unit network, are also developed and validated to compare their performance against the proposed LSTM model. The results indicated that the deep learning method RNN-LSTM has a significantly better performance (with an AUC of .70) compared to traditional classification methods such as ANN (with an AUC of 0.66) and penalized logistic regression (with an AUC of 0.63). The rationale for the better performance of the deep learning method may be due to its ability to leverage the temporal relationships of the disease state in patients over time and to capture the progression of the disease—relevant clinical information from patients' prior visits is carried forward in the memory, which may have enabled the higher predictability for the deep learning methods.

Introduction

Digitization of electronic health records (EHRs) has created momentum for a positive change in the delivery of health care. However, adoption of EHRs is only a step forward. Mere adoption of EHR did not significantly improve the quality and delivery of care. Therefore, a push has arisen to not only collect data but also to make meaningful use of it. The Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 allocates approximately $2 million to $10 million for each hospital that qualifies under “meaningful use” [23]. The goal is to improve quality and gain efficiencies in the delivery of health care. Financial incentives coupled with risk sharing models pave the way for a value-based model. In this value-based model, the value of care combines the measures of payment and patient outcomes such as rehospitalization, mortality, etc. to establish a quantitative metric of hospital performance. In this research, we focus on hospital readmission, which is defined as an event in which a patient is readmitted to the hospital for the same medical condition or for a different condition within a period of 30 days. This can be a planned or unplanned visit, and the admitting hospital can be the same hospital as the original admission or a different hospital.

Clinical decision support systems deployed on EHR databases enable real-time decision making by leveraging analytics models that can predict an event—such as rehospitalization —at the time of discharge. Systemic Lupus Erythematosus (SLE), also referred to as Lupus, is an autoimmune disease that affects multiple system organ classes. Therefore, diagnosis and disease management becomes a difficult task. Clinical manifestation of lupus is highly heterogeneous with most common organ manifestation being musculoskeletal, renal, and skin [34]. There is also a large variability in the disease severity and disease activity across patients with Lupus. Disease epidemiology shows that age-adjusted incidence rate of Lupus was 5.5 per 100,000 population and the disease prevalence was 72.8 per 100, 000 population in the US [36]. The disease prevalence across countries and regions seem to be highly variable. Asia pacific had a prevalence of 4.3–45.3 per 100, 000 [21] while the disease is rarely diagnosed in African regions. A recent study conducted by Rees et al. [35] showed that North America had the highest incidence rate of 23.2 per 100,000 person-years while Africa and Ukraine had an incidence of 0.3 per 100,000 person-years and almost negligible, close to zero, in Australia. The same study also showed that women are more frequently diagnosed with the disease than men. There was also a significant variability in the prevalence across different ethnicity groups; the disease prevalence was very high in African Americans compared to the Caucasians. This variability in the disease status, across gender, ethnicity, and regions justifies and explains the complexity in building an analytical model that can predict the clinical outcome. Research shows that identifying subgroups of patients who are high-risk and allocating resources accordingly has resulted in an absolute reduction of five readmissions per 100 index admissions in the hospital [2]. Literature suggests that high-value drivers in health care are predicting readmissions, identifying high-cost patients or patients with chronic conditions, predicting an adverse events and disease progression, and treatment optimization for conditions such as lupus [3].

Section snippets

Motivation and background

Hospital readmissions accounts for $15 billion in health care spending annually as of 2007 [33], and approximately 20% of all hospital discharges with Medicare resulted in a readmission within 30 days [31]. The Medicare Payment Advisory Commission (MedPAC) estimates that at least 12% of those hospital readmissions are avoidable. These metrics show that a large amount of health care resources are not being utilized efficiently with the fee-for-service model [7]. Although lupus is not part of the

Methodology

Deep learning is an advancement of artificial neural networks with multiple levels of representation that are created by nonlinear transformations at each level. Deep learning discovers intricate structures and complex relationships in large datasets by using algorithms that allow for tuning and readjustment of features that are used to compute the representation in a layer based on the representation from the previous layers [27]. Good features in a classification model can be learned

Results and discussion

Results from all the proposed models are shown in Table 4. Performance shows that the LSTM method had significantly higher performance compared to the traditional classification models based on the area under the curve (AUC) metric. Due to the significant imbalance in the test sample, prediction accuracy is not considered to be the best metric to gauge the model's performance, therefore AUC metric is utilized in assessing the models' performance. The AUC for both the traditional classification

Summary and conclusion

Rehospitalization is a key driver that is part of improving health care delivery, i.e., to transition from a fee-for-service model to an outcome-based model. Predicting rehospitalization helps ensure optimal utilization of the limited health care resources that are currently available, thereby improving quality as well. Results from this research show that the deep learning methodology, which can utilize longitudinal EHR data as sequential data, shows significant promise in predicting

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. This study was conducted with the data provided by, and the support from, the Center for Health Systems Innovation (CHSI) at Oklahoma State University (OSU) and the Cerner Corporation. The contents of this work are solely the responsibility of the authors and do not necessarily represent the official views of CHSI, OSU or the Cerner Corporation.

Conflict of interest statement

None Declared.

References (44)

  • N.V. Chawla et al.

    SMOTE: synthetic minority over-sampling technique

    J. Artif. Intell. Res.

    (2002)
  • E. Choi et al.

    Doctor Ai: predicting clinical events via recurrent neural networks

  • E. Choi et al.

    Using recurrent neural network models for early detection of heart failure onset

    J. Am. Med. Inf. Assoc.

    (2016)
  • R. Collobert et al.

    Natural language processing (almost) from scratch

    J. Mach. Learn. Res.

    (2011)
  • D. Delen

    Real-world Data Mining: Applied Business Analytics and Decision Making

    (2015)
  • J.P. DeShazo et al.

    A comparison of a multistate inpatient EHR database to the HCUP Nationwide Inpatient Sample

    BMC Health Serv. Res.

    (2015)
  • C.J. Edwards et al.

    Hospitalization of individuals with systemic lupus erythematosus: characteristics and predictors of outcome

    Lupus

    (2003)
  • A. Elixhauser et al.

    Readmissions to U.S. Hospitals by Diagnosis, 2010: Statistical Brief #153. United Kingdom: Europe PubMed Central

    (2006)
  • I. Goodfellow et al.

    Deep Learning

    (2016)
  • N. Hammerla et al.

    PD disease State Assessment in Naturalistic Environments Using Deep Learning. Proceedings of the 29th AAAI Conference on Artificial Intelligence

    (2015)
  • G. Hinton et al.

    Deep neural networks for acoustic modeling in speech recognition: the shared views of four research groups

    IEEE Signal Process. Mag.

    (2012)
  • S. Hochreiter et al.

    Long short-term memory

    Neural Comput.

    (1997)
  • Cited by (118)

    View all citing articles on Scopus
    View full text