Development and validation of a race-agnostic computable phenotype for kidney health in adult hospitalized patients.

Standard race adjustments for estimating glomerular filtration rate (GFR) and reference creatinine can yield a lower acute kidney injury (AKI) and chronic kidney disease (CKD) prevalence among African American patients than non-race adjusted estimates. We developed two race-agnostic computable phenotypes that assess kidney health among 139,152 subjects admitted to the University of Florida Health between 1/2012-8/2019 by removing the race modifier from the estimated GFR and estimated creatinine formula used by the race-adjusted algorithm (race-agnostic algorithm 1) and by utilizing 2021 CKD-EPI refit without race formula (race-agnostic algorithm 2) for calculations of the estimated GFR and estimated creatinine. We compared results using these algorithms to the race-adjusted algorithm in African American patients. Using clinical adjudication, we validated race-agnostic computable phenotypes developed for preadmission CKD and AKI presence on 300 cases. Race adjustment reclassified 2,113 (8%) to no CKD and 7,901 (29%) to a less severe CKD stage compared to race-agnostic algorithm 1 and reclassified 1,208 (5%) to no CKD and 4,606 (18%) to a less severe CKD stage compared to race-agnostic algorithm 2. Of 12,451 AKI encounters based on race-agnostic algorithm 1, race adjustment reclassified 591 to No AKI and 305 to a less severe AKI stage. Of 12,251 AKI encounters based on race-agnostic algorithm 2, race adjustment reclassified 382 to No AKI and 196 (1.6%) to a less severe AKI stage. The phenotyping algorithm based on refit without race formula performed well in identifying patients with CKD and AKI with a sensitivity of 100% (95% confidence interval [CI] 97%-100%) and 99% (95% CI 97%-100%) and a specificity of 88% (95% CI 82%-93%) and 98% (95% CI 93%-100%), respectively. Race-agnostic algorithms identified substantial proportions of additional patients with CKD and AKI compared to race-adjusted algorithm in African American patients. The phenotyping algorithm is promising in identifying patients with kidney disease and improving clinical decision-making.

A study of generative large language model for medical research and healthcare.

There are enormous enthusiasm and concerns in applying large language models (LLMs) to healthcare. Yet current assumptions are based on general-purpose LLMs such as ChatGPT, which are not developed for medical use. This study develops a generative clinical LLM, GatorTronGPT, using 277 billion words of text including (1) 82 billion words of clinical text from 126 clinical departments and approximately 2 million patients at the University of Florida Health and (2) 195 billion words of diverse general English text. We train GatorTronGPT using a GPT-3 architecture with up to 20 billion parameters and evaluate its utility for biomedical natural language processing (NLP) and healthcare text generation. GatorTronGPT improves biomedical natural language processing. We apply GatorTronGPT to generate 20 billion words of synthetic text. Synthetic NLP models trained using synthetic text generated by GatorTronGPT outperform models trained using real-world clinical text. Physicians' Turing test using 1 (worst) to 9 (best) scale shows that there are no significant differences in linguistic readability (p = 0.22; 6.57 of GatorTronGPT compared with 6.93 of human) and clinical relevance (p = 0.91; 7.0 of GatorTronGPT compared with 6.97 of human) and that physicians cannot differentiate them (p < 0.001). This study provides insights into the opportunities and challenges of LLMs for medical research and healthcare.

A large language model for electronic health records.

There is an increasing interest in developing artificial intelligence (AI) systems to process and interpret electronic health records (EHRs). Natural language processing (NLP) powered by pretrained language models is the key technology for medical AI systems utilizing clinical narratives. However, there are few clinical language models, the largest of which trained in the clinical domain is comparatively small at 110 million parameters (compared with billions of parameters in the general domain). It is not clear how large clinical language models with billions of parameters can help medical AI systems utilize unstructured EHRs. In this study, we develop from scratch a large clinical language model-GatorTron-using >90 billion words of text (including >82 billion words of de-identified clinical text) and systematically evaluate it on five clinical NLP tasks including clinical concept extraction, medical relation extraction, semantic textual similarity, natural language inference (NLI), and medical question answering (MQA). We examine how (1) scaling up the number of parameters and (2) scaling up the size of the training data could benefit these NLP tasks. GatorTron models scale up the clinical language model from 110 million to 8.9 billion parameters and improve five clinical NLP tasks (e.g., 9.6% and 9.5% improvement in accuracy for NLI and MQA), which can be applied to medical AI systems to improve healthcare delivery. The GatorTron models are publicly available at: https://catalog.ngc.nvidia.com/orgs/nvidia/teams/clara/models/gatortron_og .

Newly diagnosed cardiovascular disease in patients treated with immune checkpoint inhibitors: a retrospective analysis of patients at an academic tertiary care center.

BACKGROUND: Immune checkpoint inhibitors (ICIs) are a novel class of anticancer agents that have demonstrated clinical response for both solid and hematological malignancies. ICIs are associated with development of immune-related adverse events including cardiotoxicity. We estimated the incidence of newly diagnosed cardiovascular disease in patients treated with ICIs at a large, tertiary care center. METHODS: All patients with a cancer diagnosis who received any ICI treatment in the University of Florida's Integrated Data Repository from 2011 to 2017 were included. Cardiovascular disease was defined as a new ICD diagnosis code for cardiomyopathy, heart failure, arrhythmia, heart block, pericardial disease, or myocarditis after initiation of ICI treatment. RESULTS: Of 102,701 patients with a diagnosis of malignancy, 424 patients received at least one ICI. Sixty-two (14.6%) patients were diagnosed with at least one new cardiovascular disease after initiation of ICI therapy. Of the 374 patients receiving one ICI, 21 (5.6%) developed heart failure. Of the 49 patients who received two ICIs sequentially, three (6.1%) developed heart failure and/or cardiomyopathy. Incident cardiovascular disease was diagnosed at a median of 63 days after initial ICI exposure. One patient developed myocarditis 28 days after receiving nivolumab. Mortality in ICI treated patients with a concomitant diagnosis of incident cardiovascular disease was higher compared to those who did not (66.1% vs. 41.4%, odds ratio = 2.77, 1.55-4.95, p = 0.0006). CONCLUSIONS: This study suggests a high incidence of newly diagnosed cardiovascular disease after the initiation of ICI therapy in a real-world clinical setting.

Opportunity for Genotype-Guided Prescribing Among Adult Patients in 11 US Health Systems.

The value of utilizing a multigene pharmacogenetic panel to tailor pharmacotherapy is contingent on the prevalence of prescribed medications with an actionable pharmacogenetic association. The Clinical Pharmacogenetics Implementation Consortium (CPIC) has categorized over 35 gene-drug pairs as "level A," for which there is sufficiently strong evidence to recommend that genetic information be used to guide drug prescribing. The opportunity to use genetic information to tailor pharmacotherapy among adult patients was determined by elucidating the exposure to CPIC level A drugs among 11 Implementing Genomics In Practice Network (IGNITE)-affiliated health systems across the US. Inpatient and/or outpatient electronic-prescribing data were collected between January 1, 2011 and December 31, 2016 for patients ≥ 18 years of age who had at least one medical encounter that was eligible for drug prescribing in a calendar year. A median of ~ 7.2 million adult patients was available for assessment of drug prescribing per year. From 2011 to 2016, the annual estimated prevalence of exposure to at least one CPIC level A drug prescribed to unique patients ranged between 15,719 (95% confidence interval (CI): 15,658-15,781) in 2011 to 17,335 (CI: 17,283-17,386) in 2016 per 100,000 patients. The estimated annual exposure to at least 2 drugs was above 7,200 per 100,000 patients in most years of the study, reaching an apex of 7,660 (CI: 7,632-7,687) per 100,000 patients in 2014. An estimated 4,748 per 100,000 prescribing events were potentially eligible for a genotype-guided intervention. Results from this study show that a significant portion of adults treated at medical institutions across the United States is exposed to medications for which genetic information, if available, should be used to guide prescribing.

Implementation context for addressing social needs in a learning health system: a qualitative study.

INTRODUCTION: Unmet social needs contribute to growing health disparities and rising health care costs. Strategies to collect and integrate information on social needs into patients' electronic health records (EHRs) show promise for connecting patients with community resources. However, gaps remain in understanding the contextual factors that impact implementing these interventions in clinical settings. METHODS: We conducted qualitative interviews with patients and focus groups with providers (January-September 2020) in two primary care clinics to inform the implementation of a module that collects and integrates patient-reported social needs information into the EHR. Questions addressed constructs within the Theoretical Framework for Acceptability and the Consolidated Framework for Implementation Research. Data were coded deductively using team-based framework analysis, followed by inductive coding and matrix analyses. RESULTS: Forty patients participated in interviews, with 20 recruited at the clinics and 20 from home. Two focus groups were conducted with a total of 12 providers. Factors salient to acceptability and feasibility included patients' discomfort answering sensitive questions, concerns about privacy, difficulty reading/understanding module content, and technological literacy. Rapport with providers was a facilitator for patients to discuss social needs. Providers stressed that limited time with patients would be a barrier, and expressed concerns about the lack of available community resources. CONCLUSION: Findings highlight the need for flexible approaches to assessing and discussing social needs with patients. Feasibility of the intervention is contingent upon support from the health system to facilitate social needs assessment and discussion. Further study of availability of community resources is needed to ensure intervention effectiveness.

Prescribing Prevalence of Medications With Potential Genotype-Guided Dosing in Pediatric Patients.

Importance: Genotype-guided prescribing in pediatrics could prevent adverse drug reactions and improve therapeutic response. Clinical pharmacogenetic implementation guidelines are available for many medications commonly prescribed to children. Frequencies of medication prescription and actionable genotypes (genotypes where a prescribing change may be indicated) inform the potential value of pharmacogenetic implementation. Objective: To assess potential opportunities for genotype-guided prescribing in pediatric populations among multiple health systems by examining the prevalence of prescriptions for each drug with the highest level of evidence (Clinical Pharmacogenetics Implementation Consortium level A) and estimating the prevalence of potential actionable prescribing decisions. Design, Setting, and Participants: This serial cross-sectional study of prescribing prevalences in 16 health systems included electronic health records data from pediatric inpatient and outpatient encounters from January 1, 2011, to December 31, 2017. The health systems included academic medical centers with free-standing children's hospitals and community hospitals that were part of an adult health care system. Participants included approximately 2.9 million patients younger than 21 years observed per year. Data were analyzed from June 5, 2018, to April 14, 2020. Exposures: Prescription of 38 level A medications based on electronic health records. Main Outcomes and Measures: Annual prevalence of level A medication prescribing and estimated actionable exposures, calculated by combining estimated site-year prevalences across sites with each site weighted equally. Results: Data from approximately 2.9 million pediatric patients (median age, 8 [interquartile range, 2-16] years; 50.7% female, 62.3% White) were analyzed for a typical calendar year. The annual prescribing prevalence of at least 1 level A drug ranged from 7987 to 10 629 per 100 000 patients with increasing trends from 2011 to 2014. The most prescribed level A drug was the antiemetic ondansetron (annual prevalence of exposure, 8107 [95% CI, 8077-8137] per 100 000 children). Among commonly prescribed opioids, annual prevalence per 100 000 patients was 295 (95% CI, 273-317) for tramadol, 571 (95% CI, 557-586) for codeine, and 2116 (95% CI, 2097-2135) for oxycodone. The antidepressants citalopram, escitalopram, and amitriptyline were also commonly prescribed (annual prevalence, approximately 250 per 100 000 patients for each). Estimated prevalences of actionable exposures were highest for oxycodone and ondansetron (>300 per 100 000 patients annually). CYP2D6 and CYP2C19 substrates were more frequently prescribed than medications influenced by other genes. Conclusions and Relevance: These findings suggest that opportunities for pharmacogenetic implementation among pediatric patients in the US are abundant. As expected, the greatest opportunity exists with implementing CYP2D6 and CYP2C19 pharmacogenetic guidance for commonly prescribed antiemetics, analgesics, and antidepressants.

Assessing the practice of data quality evaluation in a national clinical data research network through a systematic scoping review in the era of real-world data.

OBJECTIVE: To synthesize data quality (DQ) dimensions and assessment methods of real-world data, especially electronic health records, through a systematic scoping review and to assess the practice of DQ assessment in the national Patient-centered Clinical Research Network (PCORnet). MATERIALS AND METHODS: We started with 3 widely cited DQ literature-2 reviews from Chan et al (2010) and Weiskopf et al (2013a) and 1 DQ framework from Kahn et al (2016)-and expanded our review systematically to cover relevant articles published up to February 2020. We extracted DQ dimensions and assessment methods from these studies, mapped their relationships, and organized a synthesized summarization of existing DQ dimensions and assessment methods. We reviewed the data checks employed by the PCORnet and mapped them to the synthesized DQ dimensions and methods. RESULTS: We analyzed a total of 3 reviews, 20 DQ frameworks, and 226 DQ studies and extracted 14 DQ dimensions and 10 assessment methods. We found that completeness, concordance, and correctness/accuracy were commonly assessed. Element presence, validity check, and conformance were commonly used DQ assessment methods and were the main focuses of the PCORnet data checks. DISCUSSION: Definitions of DQ dimensions and methods were not consistent in the literature, and the DQ assessment practice was not evenly distributed (eg, usability and ease-of-use were rarely discussed). Challenges in DQ assessments, given the complex and heterogeneous nature of real-world data, exist. CONCLUSION: The practice of DQ assessment is still limited in scope. Future work is warranted to generate understandable, executable, and reusable DQ measures.

Postnatal pediatric systemic antibiotic episodes during the first three years of life are not associated with mode of delivery.

BACKGROUND: Delivery by cesarean section (C-section) is associated with adverse short-term and long-term infant outcomes. Given that antibiotics during early life are prescribed for infant outcomes that are more likely among c-section deliveries, we hypothesized that postnatal antibiotic exposure will be greater among c-section infants compared to vaginally delivered infants. OBJECTIVE: The aim of this paper was to evaluate if mode of infant delivery was associated with patterns of systemic antibiotic exposure in children during their first three years. METHODS: Pediatric electronic health records from UFHealth, 2011 to 2017 were reviewed. We included singleton, term infants (37-42 weeks gestation) with a birth weight ≥ 2500 grams, with documented mode of delivery and well visits on record. Infants with a neonatal intensive care unit stay were excluded. Both oral and intravenous antibiotics for a 10-day duration were classified as a single episode. The primary outcome was antibiotic episodes in the first three years of life, and a sub-analysis was performed to compare broad-spectrum versus narrow-spectrum antibiotic exposures. RESULTS: The mean number of antibiotic episodes in 4,024 full-term infants was 0.34 (SD = 0.79) and 24.1% of infants had at least one antibiotic episode. Penicillins were the most prescribed antibiotic in children 0-1 years (66.9%) and cephalosporins were the most common antibiotic prescribed for children 1-3 years (56.2%). We did not detect a meaningful or significant rate ratio (RR) between mode of delivery and overall antibiotic episodes 1.14 (95% CI 0.99, 1.31), broad-spectrum episodes 1.19 (95% CI 0.93, 1.52, or narrow-spectrum episodes 1.14 (95% CI 0.97, 1.34). CONCLUSION: Our results do not support the hypothesis that postnatal antibiotic exposure was greater among infants delivered by cesarean section compare to infants delivered vaginally during the first three years of life.

Clinical Utility of Pharmacogene Panel-Based Testing in Patients Undergoing Percutaneous Coronary Intervention.

We aimed to estimate the utility of panel-based pharmacogenetic testing of patients undergoing percutaneous coronary intervention (PCI). Utilization of Clinical Pharmacogenetic Implementation Consortium (CPIC) level A/B drugs after PCI was estimated in a national sample of IBM MarketScan beneficiaries. Genotype data from University of Florida (UF) patients (n = 211) who underwent PCI were used to project genotype-guided opportunities among MarketScan beneficiaries with at least one (N = 105,547) and five (N = 12,462) years of follow-up data. The actual incidence of genotype-guided prescribing opportunities was determined among UF patients. In MarketScan, 50.0% (52,799/105,547) over 1 year and 68.0% (8,473/12,462) over 5 years had ≥ 1 CPIC A/B drug besides antiplatelet therapy prescribed, with a projected incidence of genotype-guided prescribing opportunities of 39% at 1 year and 52% at 5 years. Genotype-guided prescribing opportunities occurred in 32% of UF patients. Projected and actual incidence of genotype-guided opportunities among two cohorts supports the utility of panel-based testing among patients who underwent PCI.

Development and validation of an automated algorithm for identifying patients at higher risk for drug-induced acute kidney injury.

PURPOSE: Using information from institutional electronic health records, we aimed to develop dynamic predictive models to identify patients at high risk of acute kidney injury (AKI) among those who received a nephrotoxic medication during their hospital stay. METHODS: Candidate predictors were measured for each of the first 5 hospital days where a patient received a nephrotoxic medication (risk model days) to predict an AKI, using logistic regression with reduced backward variables elimination in 100 bootstrap samples. An AKI event was defined as an increase of serum creatinine ≥ 200% of a baseline SCr within 5 days after a risk model day. Final models were internally validated by replication in 100 bootstrap samples and a risk score for each patient was calculated from the validated model. As performance measures, the area under the receiver operation characteristic curves (AUC) and the number of AKI events among patients who had high risk scores were estimated. RESULTS: The study population included 62,561 admissions followed by 1,212 AKI events (1.9 events/100 admissions). We constructed 5 risk models corresponding to the first 5 hospital days where patients were exposed to at least one nephrotoxic medication. Validated AUCs of the 5 models ranged from 0.78 to 0.81. Depending on risk model day, admissions ranked in the 90th percentile of the risk score captured between 43% to 49% of all AKI events. CONCLUSION: A dynamic prediction model was built successfully for inpatient AKI with excellent discriminative validity and good calibration, allowing clinicians to focus on a select high-risk population that captures the majority of AKI events.

Development of a predictive model for drug-associated QT prolongation in the inpatient setting using electronic health record data.

PURPOSE: We aimed to construct a dynamic model for predicting severe QT interval prolongation in hospitalized patients using inpatient electronic health record (EHR) data. METHODS: A retrospective cohort consisting of all adults admitted to 2 large hospitals from January 2012 through October 2013 was established. Thirty-five risk factors for severe QT prolongation (defined as a Bazett's formula-corrected QT interval [QTc] of ≥500 msec or a QTc increase of ≥60 msec from baseline) were operationalized for automated EHR retrieval; upon univariate analyses, 26 factors were retained in models for predicting the 24-hour risk of QT events on hospital day 1 (the Day 1 model) and on hospital days 2-5 (the Days 2-5 model). RESULTS: A total of 1,672 QT prolongation events occurred over 165,847 days of risk exposure during the study period. C statistics were 0.828 for the Day 1 model and 0.813 for the Days 2-5 model. Patients in the upper 50th percentile of calculated risk scores experienced 755 of 799 QT events (94%) allocated in the Day 1 model and 804 of 873 QT events (92%) allocated in the Days 2-5 model. Among patients in the 90th percentile, the Day 1 and Days 2-5 models captured 351 of 799 (44%) and 362 of 873 (41%) QT events, respectively. CONCLUSION: The risk models derived from EHR data for all admitted patients had good predictive validity. All risk factors were operationalized from discrete EHR fields to allow full automation for real-time identification of high-risk patients. Further research to test the models in other health systems and evaluate their effectiveness on outcomes and patient care in clinical practice is recommended.

Hourly Effect of Pretreatment With IV Antibiotics on Blood Culture Positivity Rate in Emergency Department Patients.

Although it is intuitive that antibiotics administered before obtaining a blood culture would reduce the likelihood of obtaining a positive culture, it is not clear exactly how rapidly and to what extent blood becomes sterile after administration of intravenous (IV) antibiotics. Using a large data set of patients admitted from the UFHealth Shands Adult Emergency Department (ED) between 2012 and 2016 (n = 25 686), we had the opportunity to more closely examine the effect of starting IV antibiotics before vs after obtaining blood cultures. We present data on the effect of pretreatment with IV antibiotics for both septic and nonseptic ED patients on the blood culture positivity rate on an hour-by-hour basis, as well as the effects on distribution of species recovered and the impact of antibiotic resistance in empiric treatment with antibiotics.

Development and validation of a dynamic inpatient risk prediction model for clinically significant hypokalemia using electronic health record data.

Purpose: The purpose of this study was to develop a dynamic risk prediction model for inpatient hypokalemia and evaluate its predictive performance. Methods: A retrospective cohort included all admissions aged 18 years and above from 2 large tertiary hospitals in Florida over a 22-month period. Hypokalemia was defined as a potassium value of less than 3 mmol/L, and subsequent initiation of potassium supplements. Twenty-five risk factors (RF) identified from literature were operationalized using discrete electronic health record (EHR) data elements. For each of the first 5 hospital days, we modeled the probability of developing hypokalemia at the subsequent hospital day using logistic regression. Predictive performance of our model was validated with 100 bootstrap datasets and evaluated by the C statistic and Hosmer-Lemeshow goodness-of-fit test. Results: A total of 4511 hypokalemia events occurred over 263 436 hospital days (1.71%). Validated C statistics of the prediction model ranged from 0.83 (Day 1 model) to 0.86 (Day 3), while p-values for the Hosmer-Lemeshow test spanned from 0.005 (Day 1) to 0.27 (Day 4 and 5). For the Day 3 prediction, 9.9% of patients with risk scores in the 90th percentile developed hypokalemia and accounted for 60.4% of all hypokalemia events. After controlling for baseline potassium values, strong predictors included diabetic ketoacidosis, increased mineralocorticoid activity, polyuria, use of kaliuretics, use of potassium supplements and watery stool. Conclusion: This is the first risk prediction model for hypokalemia. Our model achieved excellent discrimination and adequate calibration ability. Once externally validated, this risk assessment tool could use real-time EHR information to identify individuals at the highest risk for hypokalemia and support proactive interventions by pharmacists.

MySurgeryRisk: Development and Validation of a Machine-learning Risk Algorithm for Major Complications and Death After Surgery.

OBJECTIVE: To accurately calculate the risk for postoperative complications and death after surgery in the preoperative period using machine-learning modeling of clinical data. BACKGROUND: Postoperative complications cause a 2-fold increase in the 30-day mortality and cost, and are associated with long-term consequences. The ability to precisely forecast the risk for major complications before surgery is limited. METHODS: In a single-center cohort of 51,457 surgical patients undergoing major inpatient surgery, we have developed and validated an automated analytics framework for a preoperative risk algorithm (MySurgeryRisk) that uses existing clinical data in electronic health records to forecast patient-level probabilistic risk scores for 8 major postoperative complications (acute kidney injury, sepsis, venous thromboembolism, intensive care unit admission >48 hours, mechanical ventilation >48 hours, wound, neurologic, and cardiovascular complications) and death up to 24 months after surgery. We used the area under the receiver characteristic curve (AUC) and predictiveness curves to evaluate model performance. RESULTS: MySurgeryRisk calculates probabilistic risk scores for 8 postoperative complications with AUC values ranging between 0.82 and 0.94 [99% confidence intervals (CIs) 0.81-0.94]. The model predicts the risk for death at 1, 3, 6, 12, and 24 months with AUC values ranging between 0.77 and 0.83 (99% CI 0.76-0.85). CONCLUSIONS: We constructed an automated predictive analytics framework for machine-learning algorithm with high discriminatory ability for assessing the risk of surgical complications and death using readily available preoperative electronic health records data. The feasibility of this novel algorithm implemented in real time clinical workflow requires further testing.

Implementing a hash-based privacy-preserving record linkage tool in the OneFlorida clinical research network.

OBJECTIVE: To implement an open-source tool that performs deterministic privacy-preserving record linkage (RL) in a real-world setting within a large research network. MATERIALS AND METHODS: We learned 2 efficient deterministic linkage rules using publicly available voter registration data. We then validated the 2 rules' performance with 2 manually curated gold-standard datasets linking electronic health records and claims data from 2 sources. We developed an open-source Python-based tool-OneFL Deduper-that (1) creates seeded hash codes of combinations of patients' quasi-identifiers using a cryptographic one-way hash function to achieve privacy protection and (2) links and deduplicates patient records using a central broker through matching of hash codes with a high precision and reasonable recall. RESULTS: We deployed the OneFl Deduper (https://github.com/ufbmi/onefl-deduper) in the OneFlorida, a state-based clinical research network as part of the national Patient-Centered Clinical Research Network (PCORnet). Using the gold-standard datasets, we achieved a precision of 97.25∼99.7% and a recall of 75.5%. With the tool, we deduplicated ∼3.5 million (out of ∼15 million) records down to 1.7 million unique patients across 6 health care partners and the Florida Medicaid program. We demonstrated the benefits of RL through examining different disease profiles of the linked cohorts. CONCLUSIONS: Many factors including privacy risk considerations, policies and regulations, data availability and quality, and computing resources, can impact how a RL solution is constructed in a real-world setting. Nevertheless, RL is a significant task in improving the data quality in a network so that we can draw reliable scientific discoveries from these massive data resources.

Development and validation of an automated algorithm for identifying patients at high risk for drug-induced hypoglycemia.

PURPOSE: Hypoglycemia is one of the most concerning adverse drug events in hospitalized patients. Using information from institutional electronic health records, we aimed to develop dynamic predictive models to identify patients at high risk for hypoglycemia during antihyperglycemic therapy. METHODS: The study population consisted of 21,840 patients who received antihyperglycemic medication on any of the first 5 hospital days (the "risk model days") at 2 large hospitals. Data on candidate predictors were extracted from discrete electronic health record fields to construct models for predicting hypoglycemia within 24 hours after each risk model day. Final models were internally validated by replication in 100 bootstrap samples and reapplying model parameters to the original risk population. RESULTS: The development and validation sample included 60,762 risk model days followed by 1,256 days with hypoglycemic events (2.07 events per 100 risk model days). The days 3, 4, and 5 models presented similar associations between predictors and the risk of hypoglycemia and were therefore collapsed into a single model. The strongest hypoglycemia risk factors across all 3 risk periods (day 1, day 2, and days 3-5) were blood glucose (BG) fluctuations, BG trend, history of hypoglycemia, lower body weight, lower creatinine clearance, use of long-acting or high-dose insulin, and sulfonylurea use. C statistics for the 3 models ranged from 0.844 to 0.887. Depending on the model used, risk scores in the upper 90th percentile predicted 48.5-63.1% of actual hypoglycemic events. It was estimated that by targeting only patients in the upper 90th percentile, providers would need to intervene during fewer than 9 admissions to prevent 1 hypoglycemic event. CONCLUSION: The developed prediction models were found to have excellent discriminative validity and good calibration, allowing clinicians to focus interventions on a select high-risk population in which the majority of hypoglycemic events occur.

A dynamic risk model for inpatient falls.

PURPOSE: Construction and validation of a fall risk prediction model specific to inpatients receiving fall risk-increasing drugs (FRIDs) are described. METHODS: In a retrospective cohort study of 75,036 admissions to 2 hospitals over a designated 22-month period that involved FRID exposure during the first 5 hospital days, factors influencing fall risk were investigated via logistic regression. The resultant risk prediction model was internally validated and its performance compared with that of a model based on Morse Fall Scale (MFS) scores. RESULTS: A total of 220,904 patient-days of FRID exposure were evaluated. The three most commonly administered FRIDs were oxycodone (given on 79,697 patient-days, 36.08%), morphine (52,427, 23.73%) and hydromorphone (42,063, 19.04%). Within the 90th percentile of modeled risk scores, 144 of the 466 documented falls (30.9%) were captured by the developed risk prediction model (unbiased C statistic, 0.69), as compared with 94 falls (20.2%) captured using the MFS model (unbiased C statistic, 0.62). Strong predictors of inpatient falls included a history of falling (odds ratio [OR], 1.99; 95% confidence interval (CI), 1.42-2.80); overestimation of ability to ambulate (OR, 1.53; 95% CI, 1.12-2.09); and "comorbidity predisposition," a composite measure encompassing a history of falling and 11 past diagnoses (OR, 1.60; 95% CI, 1.30-1.97). CONCLUSION: The proposed risk model for inpatient falls achieved superior predictive performance when compared with the MFS model. All risk factors were operationalized from discrete electronic health record fields, allowing full automation of real-time identification of high-risk patients.

Development and validation of a complexity score to rank hospitalized patients at risk for preventable adverse drug events.

PURPOSE: The development of risk models for 16 preventable adverse drug events (pADEs) and their aggregation into the final complexity score (C-score) are described. METHODS: Using data from 2 tertiary care facilities, logistic regression models were constructed for the first 5 hospital days that admissions were at risk for each of 16 pADEs. The best model for each pADE was validated in 100 bootstrap samples. The C-score was then aggregated and predicted individual pADE risk as the probability to develop at least 1 pADE. Using the 100 bootstrap samples for each pADE, 100 C-scores for validation were generated. RESULTS: We utilized electronic health records (EHR) data from 65,518 admissions to UF Health Shands and 18,269 admissions to UF Health Jacksonville to develop risk models for 16 pADEs. Most models had very strong discriminant validity (C-statistic > 0.8), with the highest predicted decile representing about half of manifest pADEs. Among admissions in the highest C-score decile, about two thirds experienced at least 1 pADE (C-statistic, 0.838; 95% confidence interval, 0.838-0.839). C-score precision, defined as the percentage of patients consistently (i.e., at least 95 of 100 samples) ranked in the 90th percentile, was 80-84%. CONCLUSION: The C-score was developed and validated for the identification of hospitalized patients at highest risk for pADEs. Aggregation of individual prediction models into a single score reduced its predictive power for most pADEs, compared with the individual risk models, but concentrated in the highest C-score decile a patient group more than two thirds of whom experienced at least 1 pADE.

Measurement of selected preventable adverse drug events in electronic health records: Toward developing a complexity score.

PURPOSE: The defining of a select number of high-priority preventable adverse drug events (pADEs) for measurement in the electronic health record (EHR) and the estimation of pADE incidences in two tertiary care facilities are described. METHODS: This study was part of a larger effort aimed at developing an automated electronic health record (EHR)-based complexity-score (C-score) that ranks hospitalized patients according to their risk for pADEs for clinical intervention. We developed measures for 16 high-priority pADEs often deemed preventable using discrete clinical and administrative EHR data. For each pADE we specified inclusion and exclusion criteria that were used to define risk populations for each specific pADE. The incidence of each type of pADE was then measured during a designated follow-up period considering all adult admissions to 2 large academic tertiary care hospitals, who were eligible for the pADE-specific risk populations during any of their first 5 hospital days. RESULTS: Utilizing the data from 83,787 admissions who were at risk for at least one pADE during at least one of their first five hospital days, we found that 27,193 admissions (32.5%) developed at least one pADE. Uncontrolled postsurgical pain, uncontrolled pneumonia, and drug-associated hypotension had the highest incidences with the following number of days with pADE per number of patients at risk: 13,484 of 19,640; 527 of 1,530; and 13,394 of 43,630, while drug-associated falls (446 of 75,036), drug-associated acute mental status changes (262 of 66,875) and venous thromboembolism (214 of 74,283) had the lowest incidence rates. CONCLUSION: EHR-based definitions of clinically important pADEs were developed, and the incidence of the pADEs was estimated. These definitions will be advanced for the creation of prediction models to develop a C-score for identifying patients at risk for pADEs to prioritize pharmacist intervention.