Purpose: Effectiveness of interventions to reduce hospital readmissions is limited by inadequate risk-stratification at hospital admission. The aim of this research was to develop and validate a 30-day all-cause readmission model using electronic medical records (EMR) data available within 24 hours, followed by integration of readmission risk into the electronic medical record.
Methods: We performed a retrospective cohort study among patients at Vanderbilt University Medical Center (VUMC) who were discharged alive. Patients were included if ≥ 18 years of age and admitted to a medical or surgical unit from 7/1/2009 to 6/30/2010. The outcome was readmission within 30-days from hospital discharge. 388 variables were assessed as independent predictors, obtained exclusively from electronic databases, including: demographics, admission source, number of hospital admissions in the 6 months prior, and routine laboratory tests (e.g., CBC, BMP) from the first 24 hours of admission. We developed a logistic regression model of the relationship between independent variables and all-cause 30-day readmission using modern data reduction methods. Bootstrap validation was performed with 200 replicates. We assessed discrimination and calibration with the c-statistic, Brier's score, and Hosmer-Lemeshow statistic. Finally, we tested feasibility of real-time risk calculations in the EMR.
Results: A total of 20,718 patients met the inclusion criteria, 3172 (15.3%) were readmitted to VUMC within 30 days. Overall, patients were: 53.2% male, mean age 53.5, median LOS 3.6 days (IQR 2.0 to 6.3). The final model variables included: age, emergency department admission, number of hospital admissions in the prior 6 months, hemoglobin, MCV, RDW, WBC, CO2, Cl, and BUN. The model with 10 variables had a c-statistic of 0.646 and a Brier of 0.125. The model Hosmer-Lemeshow statistic was significant (p < .0001), however this could be due to large sample size as visual calibration appeared excellent. The bootstrap validation with 200 replicates indicated minimal bias due to overfitting (slope optimism =.019). Finally, incorporation into the EMR was successfully demonstrated (See Figure).
Conclusions: Development and implementation of an all-cause real-time predictive model for 30-day hospital readmission based on data available within the first 24 hours is feasible for the entire adult hospital population. Our future work will assess whether using this model to focus interventions leads to reduced hospital readmissions.
Purpose: The Xpert MTB/RIF test enables rapid detection of tuberculosis and rifampicin resistance. The World Health Organization recommends this recently developed test for initial diagnosis in people suspected of having multi-drug resistant TB or HIV-associated-TB, and many national TB programs are moving quickly to adopt Xpert. As roll-out proceeds, it is essential to understand the potential health impact and cost-effectiveness of Xpert-based diagnostic strategies.
Method: We evaluated potential consequences of Xpert adoption in five southern African countries—Botswana, Lesotho, Namibia, South Africa, and Swaziland—where drug resistance and TB-HIV coinfection are prevalent. Analyses were conducted using a dynamic mathematical model of TB epidemiology, designed to account for the development and propagation of TB drug resistance, and the influence of epidemic HIV on TB natural history. Prior information on many TB natural history parameters is poor, and to characterize uncertainty we adopted a Bayesian estimation approach, probabilistically calibrating the model to reported data on TB prevalence, incidence, and MDR-TB prevalence by country. Using the calibrated model, we compared the status quo diagnostic algorithm, which emphasizes sputum smear, to an algorithm incorporating Xpert for initial diagnosis.
Result: Compared to status quo, implementation of Xpert would avert an estimated 132 [95% posterior interval: 55 – 284] thousand TB cases and 182 [97 – 302] thousand TB deaths in southern Africa over the 10 years following introduction, and reduce prevalence by 20-30% by 2022, with more modest reductions in incidence. Health system costs are projected to increase substantially with Xpert, requiring an additional $US 460 [294-699] million over 10 years. Antiretroviral therapy for HIV represents a substantial fraction of these additional costs, a consequence of improved survival in TB/HIV-infected populations through better TB case-finding and treatment. Relative to status quo, the Xpert strategy has an estimated cost-effectiveness of US$959 [$633-$1,485] per DALY averted over 10 years following introduction. Across the five examined countries, cost-effectiveness ratios over the same period range from $792 [$482-$1,785] in Swaziland to $1,257 [$767-$2,276] in Botswana.
Conclusion: Adoption of Xpert has potential to produce substantial changes in TB morbidity and mortality, and offers high value for money based on conventional benchmarks for cost-effectiveness in resource-limited settings. However, the additional financial burden of adoption would be substantial, including significant increases in HIV treatment costs.
Purpose: A recent study of 223,475 severely injured patients transported from the scene to trauma centers found that helicopter transport was associated with a 15% relative risk reduction in mortality compared to ground ambulance transport. In 2010, 47% of U.S. helicopter scene transports had only minor injuries. We assessed the cost-effectiveness of helicopter transport given that overtriage of patients with minor injuries to helicopter transport does not improve their outcomes.
Method: Using a Markov model, we evaluated the cost-effectiveness of helicopter scene transport relative to ground transport given triage accuracy in current practice compared with the hypothetical case of perfect triage accuracy (all patients transported have severe injury). The model followed patients from injury through prehospital care, hospitalization, first year post-discharge, and the remainder of life. Patients were trauma victims (mean age: 43; range: 18-85) with Abbreviated Injury Scores (AIS) from 1-6. Costs and survival probabilities stratified by injury severity were derived from the National Study on the Costs and Outcomes of Trauma supplemented by the National Trauma Data Bank. Transport crash risks were derived from the published literature. Outcomes included costs (2009$), quality adjusted life-years (QALYs), and incremental cost-effectiveness ratios. We used second-order Monte Carlo simulations (10,000 samples) to estimate means and confidence intervals (CI) for all outcomes.
Result: With a 15% mortality reduction and current triage accuracy, helicopter transport costs $113,306 per QALY gained (95% CI: $98,732-131,544) compared to ground ambulance transport and is never dominated or cost-saving. If triage were performed perfectly, helicopter transport would cost $67,214 per QALY gained (95% CI: $59,799-75,700), a reduction of $48,201 per QALY gained. Assuming a 15% mortality reduction, overtriage of minor injury patients would have to be reduced from 47% to 31% for helicopter transport to have at least a 95% probability of costing less than $100,000 per QALY gained. Similarly, if current triage accuracy remains the same, the mortality reduction provided by helicopter transport would need to be greater than 19%.
Conclusion: Unless overtriage of patients with minor injuries can be substantially reduced from its current level of 47%, or mortality reductions for seriously injured patients transported by helicopter are greater than was found in a recent large observational study, as currently used, helicopter scene transport is not cost-effective relative to ground transport.
Purpose:Systematic reviews that do not account for correlated outcomes may lead to biased estimates of treatment effects. We examined uncertainty in the estimate of treatment effects on two correlated outcomes in a Bayesian meta-analysis and explored how these results would alter a published cost-effectiveness analysis.
Method:We used data from a systematic review of 14 vitamin K trials that reported either bone mineral density (BMD) or fractures or both endpoints. We identified 3 trials, reporting both outcomes. We used Bayesian hierarchical random-effects meta-analysis and linear regression to sample incomplete data and model simultaneously 3 pairs of outcomes: lumbar spine BMD and all fractures; lumbar spine BMD and vertebral fractures; and, femoral neck BMD and non-vertebral fractures. We specified non-informative priors on the mean treatment effects and a Wishart prior on the inverse variance-covariance matrix. For each outcome, we estimated the population treatment effect in current trials and the predictive treatment effect in future trials. The between-study correlations and the probability that treatments jointly benefited both BMD and fractures were also calculated. We compared univariate with bivariate random-effects meta-analysis and used the population and predictive odds ratios as input parameters into a model examining the cost-effectiveness of the K vitamins for preventing fractures in women initially without osteoporosis.
Result:While the bivariate and univariate random-effects meta-analysis pooled estimates were similar, the bivariate 95% credible intervals (CrIs) were narrower and excluded implausible values. The predictive distributions shrank the most. For example, the population and predictive odds ratios for the effect of vitamin K2 on vertebral fractures and lumbar spine BMD using bivariate methods were 0.81(95% CrI: 0.5-1.1) and 0.84(95% CrI: 0.4-1.5); the corresponding univariate estimates were 0.67(95% CrI: 0.2-1.5) and 1.20(95% CrI: 0.1-5.2). The probabilities of joint benefit were 89% (vitamin K2) and 12% (vitamin K1) for vertebral fractures and lumbar spine BMD and 49% (vitamin K2) and 75% (vitamin K1) for non-vertebral fractures and femoral neck BMD. Using the results from the univariate analysis, both vitamin K2 and K1 strategies cost less than $50,000/QALY; using predictive odds ratios from the bivariate analysis, vitamin K2 strategy cost more than $100,000/QALY and vitamin K1 was cost-saving.
Conclusion:Bivariate random-effects meta-analysis can yield more plausible estimates of treatment effects that can meaningfully change the results of an economic analysis.
Purpose: Risk factors increase the incidence and severity of many chronic diseases. While some risk factors are fixed (e.g., genotypes), exposures to other risk factors (e.g., smoking) may change and are amenable to intervention. Accurate population health estimates require modeling these time-varying risk factors – a difficult task, as few longitudinal data are available. We developed a calibration procedure to infer time-varying exposures, exploiting available cross-sectional data.
Methods: We developed a simple Markov model structure that tracks the duration of continuous risk factor exposure (e.g., years as a smoker) or lack of exposure (e.g., years as a non-smoker). Risk factor exposure increases mortality risks, and exposure duration alters the probability of reducing exposure (e.g., quitting smoking); likewise, duration without exposure alters the probability of initiating exposure (e.g., starting smoking). These probabilities can vary by age and sex. The structure is deliberately simplified to facilitate incorporation into disease models (e.g., diabetes) via feasible stratifications. As an example, we calibrate sex-specific models of smoking to 10 Indian regions defined by geography and urbanicity. Indian data on sex, age, region-specific prevalence and smoking duration are derived from the Global Adult Tobacco Survey. Similarly-stratified mortality rates are derived from the Sample Registration System and age-specific smoking relative risks from the published literature. For each model, Neldor-Mead searches from 200,000 starting locations identify starting and quitting rates that minimize the difference between modeled and observed outcomes.
Results: Calibration yields close matches between modeled and observed outcomes for men and women in all regions. Generally, the probability of starting to smoke rises and falls with age (peak in teens/early 20s for men and early/mid 20s for women) while the probability of quitting smoking falls with age. Population life expectancy losses were 3-5 years for men with greater losses in higher-prevalence regions. For women, whose prevalence is 10x lower, losses were smaller. Accounting for differential starting and quitting rates based on exposure duration is potentially important as models without such variation produced greater estimates of life expectancy losses due to smoking.
Conclusions: Calibrating changes in rates of exposure for time-varying risk factors is feasible using widely-available, population-level, cross-sectional data. Incorporating exposure-change rates can improve modeled estimates of incidence and severity of related chronic diseases.