Deep Learning vs Traditional Models for Predicting Hospital Readmission among Patients with Diabetes.

A hospital readmission risk prediction tool for patients with diabetes based on electronic health record (EHR) data is needed. The optimal modeling approach, however, is unclear. In 2,836,569 encounters of 36,641 diabetes patients, deep learning (DL) long short-term memory (LSTM) models predicting unplanned, all-cause, 30-day readmission were developed and compared to several traditional models. Models used EHR data defined by a Common Data Model. The LSTM model Area Under the Receiver Operating Characteristic Curve (AUROC) was significantly greater than that of the next best traditional model [LSTM 0.79 vs Random Forest (RF) 0.72, p<0.0001]. Experiments showed that performance of the LSTM models increased as prior encounter number increased up to 30 encounters. An LSTM model with 16 selected laboratory tests yielded equivalent performance to a model with all 981 laboratory tests. This new DL model may provide the basis for a more useful readmission risk prediction tool for diabetes patients.

A case-control study of home foreclosure, health conditions, and health care utilization.

Though rates of foreclosure are at a historic high, relatively little is known about the link between foreclosure and health. We performed a case-control study to examine health conditions and health care utilization in the time period prior to foreclosure. Homeowners who received a home foreclosure notice from 2005 to 2008 were matched (by name and address) to a university hospital system in Philadelphia and compared with controls who received care from the hospital system and who lived in the same zip code as cases. Outcome measures included prevalent health conditions and visit history in the 2 years prior to foreclosure. We found that people undergoing foreclosure were similar to controls with regard to age, gender, and insurance status but significantly more likely to be African American. Rates of hypertension and renal disease were significantly higher among cases after adjustment for sociodemographic characteristics. In the 2 years prior to foreclosure, cases were more likely to visit the emergency department, have an outpatient visit, and have a no-show appointment. Cases were less likely to have a primary care physicians (PCP) visit in the 6 months immediately prior to the receipt of a foreclosure notice. The results suggest changes in health care utilization in the time period prior to foreclosure. Policies designed to decrease the incidence of home foreclosure and support people during the process should consider its association with poor health and changes in health care utilization.

Association of antibiotics in infancy with early childhood obesity.

IMPORTANCE: Obesity in children and adults is associated with significant health burdens, making prevention a public health imperative. Infancy may be a critical period when environmental factors exert a lasting effect on the risk for obesity; identifying modifiable factors may help to reduce this risk. OBJECTIVE: To assess the impact of antibiotics prescribed in infancy (ages 0-23 months) on obesity in early childhood (ages 24-59 months). DESIGN, SETTING, AND PARTICIPANTS: We conducted a cohort study spanning 2001-2013 using electronic health records. Cox proportional hazard models were used to adjust for demographic, practice, and clinical covariates. The study spanned a network of primary care practices affiliated with the Children's Hospital of Philadelphia including both teaching clinics and private practices in urban Philadelphia, Pennsylvania, and the surrounding region. All children with annual visits at ages 0 to 23 months, as well 1 or more visits at ages 24 to 59 months, were enrolled. The cohort comprised 64,580 children. EXPOSURES: Treatment episodes for prescribed antibiotics were ascertained up to 23 months of age. MAIN OUTCOMES AND MEASURES: Obesity outcomes were determined directly from anthropometric measurements using National Health and Nutrition Examination Survey 2000 body mass index norms. RESULTS: Sixty-nine percent of children were exposed to antibiotics before age 24 months, with a mean (SD) of 2.3 (1.5) episodes per child. Cumulative exposure to antibiotics was associated with later obesity (rate ratio [RR], 1.11; 95% CI, 1.02-1.21 for ≥ 4 episodes); this effect was stronger for broad-spectrum antibiotics (RR, 1.16; 95% CI, 1.06-1.29). Early exposure to broad-spectrum antibiotics was also associated with obesity (RR, 1.11; 95% CI, 1.03-1.19 at 0-5 months of age and RR, 1.09; 95% CI, 1.04-1.14 at 6-11 months of age) but narrow-spectrum drugs were not at any age or frequency. Steroid use, male sex, urban practice, public insurance, Hispanic ethnicity, and diagnosed asthma or wheezing were also predictors of obesity; common infectious diagnoses and antireflux medications were not. CONCLUSIONS AND RELEVANCE: Repeated exposure to broad-spectrum antibiotics at ages 0 to 23 months is associated with early childhood obesity. Because common childhood infections were the most frequent diagnoses co-occurring with broad-spectrum antibiotic prescription, narrowing antibiotic selection is potentially a modifiable risk factor for childhood obesity.

Using RxNorm for cross-institutional formulary data normalization within a distributed grid-computing environment.

Within the CTSA (Clinical Translational Sciences Awards) program, academic medical centers are tasked with the storage of clinical formulary data within an Integrated Data Repository (IDR) and the subsequent exposure of that data over grid computing environments for hypothesis generation and cohort selection. Formulary data collected over long periods of time across multiple institutions requires normalization of terms before those data sets can be aggregated and compared. This paper sets forth a solution to the challenge of generating derived aggregated normalized views from large, distributed data sets of clinical formulary data intended for re-use within clinical translational research.

Derivation of malignancy status from ICD-9 codes.

To assess the severity of illness of oncology patients, it is necessary to distinguish patients with a single primary tumor from patients with metastatic disease occurring at a secondary location remote from the primary site. We developed a ranked list of cancer groupings and an algorithm that could distinguish patients with primary and metastatic cancer even if no specific code for secondary cancer was recorded. In patients with metastatic disease, the algorithm should also distinguish the primary site from the secondary site.