The application of unsupervised deep learning in predictive models using electronic health records.

BACKGROUND: The main goal of this study is to explore the use of features representing patient-level electronic health record (EHR) data, generated by the unsupervised deep learning algorithm autoencoder, in predictive modeling. Since autoencoder features are unsupervised, this paper focuses on their general lower-dimensional representation of EHR information in a wide variety of predictive tasks. METHODS: We compare the model with autoencoder features to traditional models: logistic model with least absolute shrinkage and selection operator (LASSO) and Random Forest algorithm. In addition, we include a predictive model using a small subset of response-specific variables (Simple Reg) and a model combining these variables with features from autoencoder (Enhanced Reg). We performed the study first on simulated data that mimics real world EHR data and then on actual EHR data from eight Advocate hospitals. RESULTS: On simulated data with incorrect categories and missing data, the precision for autoencoder is 24.16% when fixing recall at 0.7, which is higher than Random Forest (23.61%) and lower than LASSO (25.32%). The precision is 20.92% in Simple Reg and improves to 24.89% in Enhanced Reg. When using real EHR data to predict the 30-day readmission rate, the precision of autoencoder is 19.04%, which again is higher than Random Forest (18.48%) and lower than LASSO (19.70%). The precisions for Simple Reg and Enhanced Reg are 18.70 and 19.69% respectively. That is, Enhanced Reg can have competitive prediction performance compared to LASSO. In addition, results show that Enhanced Reg usually relies on fewer features under the setting of simulations of this paper. CONCLUSIONS: We conclude that autoencoder can create useful features representing the entire space of EHR data and which are applicable to a wide array of predictive tasks. Together with important response-specific predictors, we can derive efficient and robust predictive models with less labor in data extraction and model training.

The association between outpatient follow-up visits and all-cause non-elective 30-day readmissions: A retrospective observational cohort study.

BACKGROUND: As an effort to reduce hospital readmissions, early follow-up visits were recommended by the Society of Hospital Medicine. However, published literature on the effect of follow-up visits is limited with mixed conclusions. Our goal here is to fully explore the relationship between follow-up visits and the all-cause non-elective 30-day readmission rate (RR) after adjusting for confounders. METHODS AND RESULTS: To conduct this retrospective observational study, we extracted data for 55,378 adult inpatients from Advocate Health Care, a large, multi-hospital system serving a diverse population in a major metropolitan area. These patients were discharged to Home or Home with Home Health services between June 1, 2013 and April 30, 2015. Our findings from time-dependent Cox proportional hazard models showed that follow-up visits were significantly associated with a reduced RR (adjusted hazard ratio: 0.86; 95% CI: 0.82-0.91), but in a complicated way because the interaction between follow-up visits and a readmission risk score was significant with p-value < 0.001. Our analysis using logistic models on an adjusted data set confirmed the above findings with the following additional results. First, time matter. Follow-up visits within 2 days were associated with the greatest reduction in RR (adjusted odds ratio: 0.72; 95% CI: 0.63-0.83). Visits beyond 2 days were also associated with a reduction in RR, but the strength of the effect decreased as the time between discharge and follow-up visit increased. Second, the strength of such association varied for patients with different readmission risk scores. Patients with a risk score of 0.113, high but not extremely high risk, had the greatest reduction in RR from follow-up visits. Patients with an extremely high risk score (> 0.334) saw no RR reduction from follow-up visits. Third, a patient was much more likely to have a 2-day follow-up visit if that visit was scheduled before the patient was discharged from the hospital (30% versus < 5%). CONCLUSIONS: Follow-up visits are associated with a reduction in readmission risk. The timing of follow-up visits can be important: beyond two days, the earlier, the better. The effect of follow-up visits is more significant for patients with a high but not extremely high risk of readmission.

Comparison of predictive modeling approaches for 30-day all-cause non-elective readmission risk.

BACKGROUND: This paper explores the importance of electronic medical records (EMR) for predicting 30-day all-cause non-elective readmission risk of patients and presents a comparison of prediction performance of commonly used methods. METHODS: The data are extracted from eight Advocate Health Care hospitals. Index admissions are excluded from the cohort if they are observation, inpatient admissions for psychiatry, skilled nursing, hospice, rehabilitation, maternal and newborn visits, or if the patient expires during the index admission. Data are randomly and repeatedly divided into fitting and validating sets for cross validations. Approaches including LACE, STEPWISE logistic, LASSO logistic, and AdaBoost, are compared with sample sizes varying from 2,500 to 80,000. RESULTS: Our results confirm that LACE has moderate discrimination power with the area under receiver operating characteristic curve (AUC) around 0.65-0.66, which can be improved to 0.73-0.74 when additional variables from EMR are considered. These variables include Inpatient in the last six months, Number of emergency room visits or inpatients in the last year, Braden score, Polypharmacy, Employment status, Discharge disposition, Albumin level, and medical condition variables such as Leukemia, Malignancy, Renal failure with hemodialysis, History of alcohol substance abuse, Dementia and Trauma. When sample size is small (≤5000), LASSO is the best; when sample size is large (≥20,000), the predictive performance is similar. The STEPWISE method has a slightly lower AUC (0.734) comparing to LASSO (0.737) and AdaBoost (0.737). More than one half of the selected predictors can be false positives when using a single method and a single division of fitting/validating data. CONCLUSIONS: True predictors can be identified by repeatedly dividing data into fitting/validating subsets and referring the final model based on summarizing results. LASSO is a better alternative to the STEPWISE logistic regression, especially when sample size is not large. The evidence for adequate sample size can be explored by fitting models on gradually reduced samples. Our model comparison strategy is not only good for 30-day all-cause non-elective readmission risk predictions, but also applicable to other types of predictive models in clinical studies.

Evaluation of International Classification of Diseases, Ninth Revision, Clinical Modification Codes for reporting methicillin-resistant Staphylococcus aureus infections at a hospital in Illinois.

BACKGROUND: States, including Illinois, have passed legislation mandating the use of International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes for reporting healthcare-associated infections, such as methicillin-resistant Staphylococcus aureus (MRSA). OBJECTIVE: To evaluate the sensitivity of ICD-9-CM code combinations for detection of MRSA infection and to understand implications for reporting. METHODS: We reviewed discharge and microbiology databases from July through August of 2005, 2006, and 2007 for ICD-9-CM codes or microbiology results suggesting MRSA infection at a tertiary care hospital near Chicago, Illinois. Medical records were reviewed to confirm MRSA infection. Time from admission to first positive MRSA culture result was evaluated to identify hospital-onset MRSA (HO-MRSA) infections. The sensitivity of MRSA code combinations for detecting confirmed MRSA infections was calculated using all codes present in the discharge record (up to 15); the effect of reviewing only 9 diagnosis codes, the number reported to the Centers for Medicare and Medicaid Services, was also evaluated. The sensitivity of the combination of diagnosis codes for detection of HO-MRSA infections was compared with that for community-onset MRSA (CO-MRSA) infections. RESULTS: We identified 571 potential MRSA infections with the use of screening criteria; 403 (71%) were confirmed MRSA infections, of which 61 (15%) were classified as HO-MRSA. The sensitivity of MRSA code combinations was 59% for all confirmed MRSA infections when 15 diagnoses were reviewed compared with 31% if only 9 diagnoses were reviewed (P < .001). The sensitivity of code combinations was 33% for HO-MRSA infections compared with 62% for CO-MRSA infections (P < .001). CONCLUSIONS: Limiting analysis to 9 diagnosis codes resulted in low sensitivity. Furthermore, code combinations were better at revealing CO-MRSA infections than HO-MRSA infections. These limitations could compromise the validity of ICD-9-CM codes for interfacility comparisons and for reporting of healthcare-associated MRSA infections.