Exploiting hierarchy in medical concept embedding*

Objective To construct and publicly release a set of medical concept embeddings for codes following the ICD-10 coding standard which explicitly incorporate hierarchical information from medical codes into the embedding formulation. Materials and Methods We trained concept embeddings using several new extensions to the Word2Vec algorithm using a dataset of approximately 600,000 patients from a major integrated healthcare organization in the Mid-Atlantic US. Our concept embeddings included additional entities to account for the medical categories assigned to codes by the Clinical Classification Software Revised (CCSR) dataset. We compare these results to sets of publicly released pretrained embeddings and alternative training methodologies. Results We found that Word2Vec models which included hierarchical data outperformed ordinary Word2Vec alternatives on tasks which compared naïve clusters to canonical ones provided by CCSR. Our Skip-Gram model with both codes and categories achieved 61.4% normalized mutual information with canonical labels in comparison to 57.5% with traditional Skip-Gram. In models operating on two different outcomes, we found that including hierarchical embedding data improved classification performance 96.2% of the time. When controlling for all other variables, we found that co-training embeddings improved classification performance 66.7% of the time. We found that all models outperformed our competitive benchmarks. Discussion We found significant evidence that our proposed algorithms can express the hierarchical structure of medical codes more fully than ordinary Word2Vec models, and that this improvement carries forward into classification tasks. As part of this publication, we have released several sets of pretrained medical concept embeddings using the ICD-10 standard which significantly outperform other well-known pretrained vectors on our tested outcomes.

Use of clinical classification software to differentiate good versus poor prognostic patients with ovarian cancer: A SEER-Medicare database analysis from the viera study.

e18069 Background: Historically, recurrence in ovarian cancer (OC) following first-line (1L) chemotherapy (CT) occurs in up to 80% of patients within 2 years. The Clinical Classification Software (CCS) systematically classifies thousands of ICD-9 codes into a smaller number of clinically meaningful categories. We sought to use CCS and other routinely collected variables to differentiate the clinical and demographic profiles of patients with good prognosis (GP) versus poor prognosis (PP) in the United States (US). Methods: This was a retrospective cohort study of newly diagnosed (FIGO stage II - IV), treatment-naïve patients, ≥ 66 years, who received 4-10 cycles of platinum-based 1L CT between Jan 2009 - Dec 2015 using the SEER-Medicare database, a nationally representative cancer registry. Patient were assumed to have progressed to a subsequent line of therapy following a gap between consecutive CT cycles ≥ 63 days. Patients were classified as GP if alive ≥4 years with no further treatment following 1L CT; PP was defined as receipt of ≥2L CT within 12 months of initial 1L CT. Demographic and prognostic characteristics were assessed during a 6-month baseline period prior to initiation of 1L CT. We assessed clinically meaningful differences in baseline characteristics with absolute standardized differences (ASD) using a threshold of 0.1 (indicating negligible difference between two cohorts). Results: There were a total of 2,262 patients (mean age: 74.6 ±6.2 years) including 251 GP (11%) and 209 PP (9%) patients (table below). PP patients were significantly more likely to be older than 70 years, and present at stage IV, liver disease and ascites, and anemia at diagnosis. PP patients were also less likely to have primary debulking surgery. Conclusions: Approximately one tenth of OC patients received no further treatment 4 years after the initial treatment with contemporary standard of care. GP may be differentiated from PP on the basis of commonly used clinical characteristics such as stage and also specific comorbidities such as liver disease and ascites. [Table: see text]

MEDICAL COMPLICATIONS AND INJURY LEADING TO EMERGENCY DEPARTMENT USE AMONG OLDER ADULTS

Medical injury consistently ranks among the most expensive hospital stay diagnoses and represents a frequent cause of hospital readmission. Although older adults are at greater risk of medical injury, in part, because of greater incidence of comorbidity and increased medical complexity, little is known about the burden of medical injury leading to ED use or the costs and outcomes associated with these events among older adults. In response, this study used nationally representative data from the 2014 Nationwide Emergency Department Survey to examine the epidemiology of older adult ED-visits for medical injury. Principal diagnosis codes were grouped using AHRQ’s Clinical Classification Software to identify medical injury-related visits. Results indicated that in 2014, 506,466 ED-visits for medical injuries occurred, comprising 2% of all older adult ED-visits. Leading causes of medical injury included malfunction of device, implant and grafts (24%); infection and inflammation of internal prosthetic device, implant, and graft (16%), and other complications of surgical and medical procedures (15%). Risk factors for medical injury included being male, Medicaid as primary payor, and number of chronic conditions. Multinominal logistic regression and multivariate regression results indicate that Medical injury-related ED visits were associated with higher hospitalization risk (RRR=2.08, p<0.000), 27% longer hospital stays, and 24% higher total charges relative to non-medical injury related visits. However, medical injury was not associated with risk of death after adjustment. Study findings suggest that ED-visits for medical injury occur frequently among older adults and are associated with significant burden and cost.

2677. Infection-Related Outcomes in Patients With Malignancy-Related Febrile Neutropenia: A National Perspective

Background Febrile neutropenia (FBN) is a life-threatening oncological emergency requiring hospitalization and early treatment with broad-spectrum antibiotics. We aimed to study differences in infection-related outcomes for febrile neutropenia in various malignancies. Methods The National Inpatient Sample (NIS) data set was queried from 2007 to 2014 to identify all patients with a diagnosis of neutropenic fever (ICD-9: 780.6x and 288.5x or 288.0x or 284.1x). Diagnoses for various cancers were determined via their respective Clinical Classification Software (CCS) codes. Diagnoses of pneumonia (481.x, 482x), bacterial meningitis (320.x), Clostridium difficile (008.45), infectious colitis due to neoplastic agents (009.x), urinary tract infection (599.0x), pyelonephritis (590.1x, 590.80), skin and soft-tissue infection (682.x, 684.x, 686.8x, 686.9x), mucositis (528.01), influenza (CCS 487), sepsis (995.91), severe sepsis (995.92), septic shock (785.52), E. coli septicemia (038.42), Pseudomonas septicemia (038.43), MRSA septicemia (038.12) and Streptococcal septicemia (038.0) were identified using their respective ICD/CCS codes. Variables were analyzed via multivariate analysis using the program SAS. Results We studied 381,043 patients with FBN. Leukemia was the most common malignancy associated with FBN (140,190, patients, 36.8%). Meningitis was found to be significantly associated with brain cancer, while other infections were associated with a range of malignancies. (Table.1) Methicillin-resistant Staphylococcus aureus was associated with cancers of the bone, breast, uterus and non- hodgkins lymphoma, while other microorganisms varied across different malignancies (Table 2). Septic Shock was associated with cancer of the pancreas, lung, bone, breast, leukemia, bladder, kidney, thyroid, myeloma, prostate, testis, cervix, brain, melanoma, non-hodgkins lymphoma, compared with other malignancies (Table 3). Conclusion Pathogen-specific and targeted antibiotic therapy is the cornerstone of treatment in FBN. Our study provides evidence of specific presentations and organisms causing infections in various malignancies. We hope that further outcomes-based research will provide objective evidence of certain high-risk infections, improving patient outcomes and minimizing redundant testing. Disclosures All authors: No reported disclosures.

Comparison of demographic variables related to 30-day readmission in cancers with similar risk factors: Analysis of 290,270 hospitalizations.

e15596 Background: There has been increasing focus on reducing hospital readmissions which are an ongoing economic challenge. This study was targeted at cancer of the esophagus (CE), stomach (CS), and pancreas (CP) which share similar risk factors to identify if these patient subsets also share risk for hospital readmission. Methods: Data was extracted from Nationwide Inpatient Sample database, which approximates 20% of U.S. community hospitals. Patients with a diagnosis CE, CS, and CP were identified by Clinical Classification Software code 11, 12, and 17, respectively. Demographic parameters associated with high readmission rates were collected for patients readmitted within 30 days over 2009-2013. Chi-square test was used to asses differences between variables. Results: We identified a total of 290,270 hospitalizations over the 5-year period of which 26.2% were readmitted within 30 days. The total number of readmissions were comparable for CS (28.0%), CP (27.89%), and CE (26.81%). Patients 45 to 64 years (28.58%, p < 0.001), Medicaid insurances (33.4%, p < 0.001), and metropolitan areas (26.8%, p < 0.001) were associated with higher 30-day readmission rates amongst all three cancers. Males had higher readmission rates in CS (25.64%, p < 0.001) and CP (28.6%, p < 0.001) compared to females in CE (28.0%, p < 0.001). Lowest median income for zip code patients were readmitted more often in CS (26.2%, p < 0.001) vs. third median income in CP (27.78, p < 0.001) and fourth (highest) median income in CE (28.38 %, p < 0.001). The most common identified causes of readmission in all three cancers were complications of surgical procedures or medical care (8.51% ± 2.03%), and septicemia (6.5% ± 1.95%). Conclusions: Our analysis shows that some variables for readmission are similar however more studies are needed to further elucidate whether this is primarily due to analogous treatment techniques. If so interventions to limit readmissions related to these treatments should be taken. Differences in readmissions rate amongst income classes could be related to the ease of accessing the most endorsed treatments for each individual cancer, however further studies are needed to clarify this discrepancy.

Characterizing the costs of complications after cystectomy: Can we target the primary drivers?

304 Background: Radical Cystectomy (RC) is subject to substantial morbidity and patients face complication rates as high as 64% at 90-days. Understanding the costs of complications after RC is essential to improving care. We studied the financial cost of different categories of complications after RC in order to identify drivers of expenditures. Methods: Using the Premier Hospital Database we identified adult patients who underwent RC for bladder cancer from 600 hospitals across the US from 2003-2013. Ninety-day complications were captured using ICD9 codes. Complications were categorized according to Agency for Healthcare Research and Quality Clinical Classification Software. The primary outcome was cost of complication and secondary outcomes were mortality, length of stay (LOS), and discharge disposition. A generalized liner model conforming to a gamma distribution was used to evaluate cost data. Analyses were survey weighted, and all models were adjusted for patient (age, race, obesity, marital status, payer), hospital (bed size, teaching affiliation, rural, region), and surgery characteristics (lymphadenectomy, continent diversion, robotic, operative time, transfusion, surgeon volume, hospital volume) and clustered by hospital. Results: We identified 9,137 RC patients, representing a weighting population of 57,553 patients. The top four most costly complications were venous thromboembolism (VTE $17547), soft tissue ($13523), gastrointestinal (GI $8663), and infectious (non-wound, i.e. sepsis, $7930). Pharmacy related costs were the primary driver of VTE costs. LOS was increased in each complication by 1.7 days for infectious, 4.5 days for soft tissue, 3.5 days for GI, and 3 days for VTE. Being married, having fewer comorbidities, larger hospitals, teaching hospitals, shorter operations, lack of transfusions, high volume hospitals, and high volume surgeons were associated with statistically significantly lower costs of complications after cystectomy. Conclusions: VTE, soft tissue, and GI complications are the most expensive complications after cystectomy, and thereby highlight potential candidates for future quality improvement initiatives.

Abstract TP402: Sickle Cell Disease in Pediatric Stroke - Inpatient Resource Utilization in the Decade Following STOP

Background: Sickle cell disease (SCD) is an important risk factor for stroke in children. Natural history studies demonstrate that greater than 10% of hemoglobin SS patients suffered ischemic stroke prior to age 20 years. In 1998, the Stroke Prevention Trial in Sickle Cell Anemia (STOP) successfully demonstrated the role for routine transfusion therapy in reducing stroke in at risk SCD patients. Fullerton and colleagues then found that first time stroke in SCD decreased in Californian children in the 2 years following STOP. We investigated the stroke rate and health care utilization of children with SCD for two calendar years in the decade following publication of the STOP trial using a national inpatient database. Methods: The 2000 and 2009 Kids’ Inpatient Database (KID) were used for analysis. SCD and stroke cases were identified by ICD-9 codes 282.6x, 430, 431, 432.9, 434.X1, 434.9, 435.9. We queried the KID procedural clinical classification software for utilization of services pertinent to SCD and stroke; transfusion, MRI, and cerebral angio. Results: In 2000, SCD was a discharge diagnosis in 34,294 children and 158 (0.46%) children had SCD and stroke. By 2009, discharges with SCD rose to 37,082 children with 212 (0.57%) children carrying both diagnoses. In 2000 and 2009, AIS is the most common stroke type at 83%, males account for 53% of stroke and black race was reported by 92% of SCD and stroke subjects. Procedure utilization is higher in the SCD and stroke population than in SCD without stroke (Figure 1). Blood transfusion is the most common procedure in both study years, significantly higher in stroke subjects. Conclusion: For pediatric inpatients with SCD, blood transfusion and diagnostic cerebrovascular procedures were significantly more common in the cohort with comorbid stroke. In the decade after STOP, children hospitalized with SCD and stroke represented less than 0.6% of the total inpatient SCD population.

Using the Electronic Health Record to Identify Educational Gaps for Internal Medicine Interns

ABSTRACT Background An important component of internal medicine residency is clinical immersion in core rotations to expose first-year residents to common diagnoses. Objective Quantify intern experience with common diagnoses through clinical documentation in an electronic health record. Methods We analyzed all clinical notes written by postgraduate year (PGY) 1, PGY-2, and PGY-3 residents on medicine service at an academic medical center July 1, 2012, through June 30, 2014. We quantified the number of notes written by PGY-1s at 1 of 3 hospitals where they rotate, by the number of notes written about patients with a specific principal billing diagnosis, which we defined as diagnosis-days. We used the International Classification of Diseases 9 (ICD-9) and the Clinical Classification Software (CCS) to group the diagnoses. Results We analyzed 53 066 clinical notes covering 10 022 hospitalizations with 1436 different ICD-9 diagnoses spanning 217 CCS diagnostic categories. The 10 most common ICD-9 diagnoses accounted for 23% of diagnosis-days, while the 10 most common CCS groupings accounted for more than 40% of the diagnosis-days. Of 122 PGY-1s, 107 (88%) spent at least 2 months on the service, and 3% were exposed to all of the top 10 ICD-9 diagnoses, while 31% had experience with fewer than 5 of the top 10 diagnoses. In addition, 17% of PGY-1s saw all top 10 CCS diagnoses, and 5% had exposure to fewer than 5 CCS diagnoses. Conclusions Automated detection of clinical experience may help programs review inpatient clinical experiences of PGY-1s.

Abstract WP189: Trends in Comorbidities Among Stroke-related Hospitalizations in California, 2000-2013

Background: Preventive care for adults with hypertension and diabetes has improved over the recent decade. This study aims to identify trends of these and other comorbidities for patients with stroke-related hospitalizations in California from 2000 to 2013. Methods: This retrospective study is based on patient discharge information compiled by the California Office of Statewide Health and Planning Development from 2000 to 2013. The study population is comprised of residents, and represents over 31.68 million hospital events. Stroke hospitalizations were identified using the first listed diagnoses in the hospital records, ICD-9-CM codes 430 to 436. Comorbidities were identified using ICD-9-CM codes in 24 other diagnostic fields in each stroke hospitalization, aided by AHRQ’s Clinical Classification Software. Cochran Armitage trend test was employed to identify comorbidity trends from 2000 to 2013. Stroke hospitalizations were adjusted for sex, age, race/ethnicity and insurance type using multivariable logistic regression for each of the comorbidities analyzed. Results: Over 1 million hospitalizations and 77,908 in-hospital mortalities for stroke were identified from 2000 to 2013. Disorders of lipid metabolism among stroke-related hospitalizations rose by 230.5% from 16.7% in 2000 to 55.1% in 2013. Diabetes mellitus increased by 45.1% from 27.0% in 2000 to 39.2% in 2013. Cardiac dysrhythmia rose by 36.6% from 21.0% in 2000 to 28.7% in 2013 while hypertension rose by 22.0% from 67.7% in 2000 to 82.6% in 2013. The trend test showed significant upward trends (p <0.001) for these four comorbidities during the study period. After adjusting for patient demographics and insurance type, all four comorbidities had a positive association with stroke-related hospitalizations: diabetes mellitus (AOR 1.247; 95% CI=1.228-1.267), hypertension (AOR 2.685;[95% CI=2.631-2.740), disorders of lipid metabolism (AOR 2.095; 95% CI=2.063-2.127) and cardiac dysrhythmias (AOR 1.193; 95% CI=1.172-1.213). Conclusion: Diabetes mellitus, hypertension, disorders of lipid metabolism and cardiac dysrhythmia are positively associated with stoke-related hospitalizations.