Surgical scheduling via optimization and machine learning with long-tailed data : Health care management science, in press.

Using data from cardiovascular surgery patients with long and highly variable post-surgical lengths of stay (LOS), we develop a modeling framework to reduce recovery unit congestion. We estimate the LOS and its probability distribution using machine learning models, schedule procedures on a rolling basis using a variety of optimization models, and estimate performance with simulation. The machine learning models achieved only modest LOS prediction accuracy, despite access to a very rich set of patient characteristics. Compared to the current paper-based system used in the hospital, most optimization models failed to reduce congestion without increasing wait times for surgery. A conservative stochastic optimization with sufficient sampling to capture the long tail of the LOS distribution outperformed the current manual process and other stochastic and robust optimization approaches. These results highlight the perils of using oversimplified distributional models of LOS for scheduling procedures and the importance of using optimization methods well-suited to dealing with long-tailed behavior.

A quantitative model to ensure capacity sufficient for timely access to care in a remote patient monitoring program.

INTRODUCTION: Algorithm-enabled remote patient monitoring (RPM) programs pose novel operational challenges. For clinics developing and deploying such programs, no standardized model is available to ensure capacity sufficient for timely access to care. We developed a flexible model and interactive dashboard of capacity planning for whole-population RPM-based care for T1D. METHODS: Data were gathered from a weekly RPM program for 277 paediatric patients with T1D at a paediatric academic medical centre. Through the analysis of 2 years of observational operational data and iterative interviews with the care team, we identified the primary operational, population, and workforce metrics that drive demand for care providers. Based on these metrics, an interactive model was designed to facilitate capacity planning and deployed as a dashboard. RESULTS: The primary population-level drivers of demand are the number of patients in the program, the rate at which patients enrol and graduate from the program, and the average frequency at which patients require a review of their data. The primary modifiable clinic-level drivers of capacity are the number of care providers, the time required to review patient data and contact a patient, and the number of hours each provider allocates to the program each week. At the institution studied, the model identified a variety of practical operational approaches to better match the demand for patient care. CONCLUSION: We designed a generalizable, systematic model for capacity planning for a paediatric endocrinology clinic providing RPM for T1D. We deployed this model as an interactive dashboard and used it to facilitate expansion of a novel care program (4 T Study) for newly diagnosed patients with T1D. This model may facilitate the systematic design of RPM-based care programs.

A model to design financially sustainable algorithm-enabled remote patient monitoring for pediatric type 1 diabetes care.

Introduction: Population-level algorithm-enabled remote patient monitoring (RPM) based on continuous glucose monitor (CGM) data review has been shown to improve clinical outcomes in diabetes patients, especially children. However, existing reimbursement models are geared towards the direct provision of clinic care, not population health management. We developed a financial model to assist pediatric type 1 diabetes (T1D) clinics design financially sustainable RPM programs based on algorithm-enabled review of CGM data. Methods: Data were gathered from a weekly RPM program for 302 pediatric patients with T1D at Lucile Packard Children's Hospital. We created a customizable financial model to calculate the yearly marginal costs and revenues of providing diabetes education. We consider a baseline or status quo scenario and compare it to two different care delivery scenarios, in which routine appointments are supplemented with algorithm-enabled, flexible, message-based contacts delivered according to patient need. We use the model to estimate the minimum reimbursement rate needed for telemedicine contacts to maintain revenue-neutrality and not suffer an adverse impact to the bottom line. Results: The financial model estimates that in both scenarios, an average reimbursement rate of roughly $10.00 USD per telehealth interaction would be sufficient to maintain revenue-neutrality. Algorithm-enabled RPM could potentially be billed for using existing RPM CPT codes and lead to margin expansion. Conclusion: We designed a model which evaluates the financial impact of adopting algorithm-enabled RPM in a pediatric endocrinology clinic serving T1D patients. This model establishes a clear threshold reimbursement value for maintaining revenue-neutrality, as well as an estimate of potential RPM reimbursement revenue which could be billed for. It may serve as a useful financial-planning tool for a pediatric T1D clinic seeking to leverage algorithm-enabled RPM to provide flexible, more timely interventions to its patients.

Assessment of physician training and prediction of workforce needs in paediatric cardiac intensive care in the United States.

OBJECTIVE: To assess the training and the future workforce needs of paediatric cardiac critical care faculty. DESIGN: REDCap surveys were sent May-August 2019 to medical directors and faculty at the 120 US centres participating in the Society of Thoracic Surgeons Congenital Heart Surgery Database. Faculty and directors were asked about personal training pathway and planned employment changes. Directors were additionally asked for current faculty numbers, expected job openings, presence of training programmes, and numbers of trainees. Predictive modelling of the workforce was performed using respondents' data. Patient volume was projected from US Census data and compared to projected provider availability. MEASUREMENTS AND MAIN RESULTS: Sixty-six per cent (79/120) of directors and 62% (294/477) of contacted faculty responded. Most respondents had training that incorporated critical care medicine with the majority completing training beyond categorical fellowship. Younger respondents and those in dedicated cardiac ICUs were more significantly likely to have advanced training or dual fellowships in cardiology and critical care medicine. An estimated 49-63 faculty enter the workforce annually from various training pathways. Based on modelling, these faculty will likely fill current and projected open positions over the next 5 years. CONCLUSIONS: Paediatric cardiac critical care training has evolved, such that the majority of faculty now have dual fellowship or advanced training. The projected number of incoming faculty will likely fill open positions within the next 5 years. Institutions with existing or anticipated training programmes should be cognisant of these data and prepare graduates for an increasingly competitive market.

Population-level management of type 1 diabetes via continuous glucose monitoring and algorithm-enabled patient prioritization: Precision health meets population health.

OBJECTIVE: To develop and scale algorithm-enabled patient prioritization to improve population-level management of type 1 diabetes (T1D) in a pediatric clinic with fixed resources, using telemedicine and remote monitoring of patients via continuous glucose monitor (CGM) data review. RESEARCH DESIGN AND METHODS: We adapted consensus glucose targets for T1D patients using CGM to identify interpretable clinical criteria to prioritize patients for weekly provider review. The criteria were constructed to manage the number of patients reviewed weekly and identify patients who most needed provider contact. We developed an interactive dashboard to display CGM data relevant for the patients prioritized for review. RESULTS: The introduction of the new criteria and interactive dashboard was associated with a 60% reduction in the mean time spent by diabetes team members who remotely and asynchronously reviewed patient data and contacted patients, from 3.2 ± 0.20 to 1.3 ± 0.24 min per patient per week. Given fixed resources for review, this corresponded to an estimated 147% increase in weekly clinic capacity. Patients who qualified for and received remote review (n = 58) have associated 8.8 percentage points (pp) (95% CI = 0.6-16.9 pp) greater time-in-range (70-180 mg/dl) glucoses compared to 25 control patients who did not qualify at 12 months after T1D onset. CONCLUSIONS: An algorithm-enabled prioritization of T1D patients with CGM for asynchronous remote review reduced provider time spent per patient and was associated with improved time-in-range.

Improved individual and population-level HbA1c estimation using CGM data and patient characteristics.

Machine learning and linear regression models using CGM and participant data reduced HbA1c estimation error by up to 26% compared to the GMI formula, and exhibit superior performance in estimating the median of HbA1c at the cohort level, potentially of value for remote clinical trials interrupted by COVID-19.

The development and deployment of a model for hospital-level COVID-19 associated patient demand intervals from consistent estimators (DICE).

Hospitals commonly project demand for their services by combining their historical share of regional demand with forecasts of total regional demand. Hospital-specific forecasts of demand that provide prediction intervals, rather than point estimates, may facilitate better managerial decisions, especially when demand overage and underage are associated with high, asymmetric costs. Regional point forecasts of patient demand are commonly available, e.g., for the number of people requiring hospitalization due to an epidemic such as COVID-19. However, even in this common setting, no probabilistic, consistent, computationally tractable forecast is available for the fraction of patients in a region that a particular institution should expect. We introduce such a forecast, DICE (Demand Intervals from Consistent Estimators). We describe its development and deployment at an academic medical center in California during the 'second wave' of COVID-19 in the Unite States. We show that DICE is consistent under mild assumptions and suitable for use with perfect, biased and unbiased regional forecasts. We evaluate its performance on empirical data from a large academic medical center as well as on synthetic data.

Reducing administrative costs in US health care: Assessing single payer and its alternatives.

OBJECTIVE: Excess administrative costs in the US health care system are routinely referenced as a justification for comprehensive reform. While there is agreement that these costs are too high, there is little understanding of what generates administrative costs and what policy options might mitigate them. DATA SOURCES: Literature review and national utilization and expenditure data. STUDY DESIGN: We developed a simulation model of physician billing and insurance-related (BIR) costs to estimate how certain policy reforms would generate savings. Our model is based on structural elements of the payment process in the United States and considers each provider's number of health plan contracts, the number of features in each health plan, the clinical and nonclinical processes required to submit a bill for payment, and the compliance costs associated with medical billing. DATA EXTRACTION: For several types of visits, we estimated fixed and variable costs of the billing process. We used the model to estimate the BIR costs at a national level under a variety of policy scenarios, including variations of a single payer "Medicare-for-All" model that extends fee-for-service Medicare to the entire population and policy efforts to reduce administrative costs in a multi-payer model. We conducted sensitivity analyses of a wide variety of model parameters. PRINCIPAL FINDINGS: Our model estimates that national BIR costs are reduced between 33% and 53% in Medicare-for-All style single-payer models and between 27% and 63% in various multi-payer models. Under a wide range of assumptions and sensitivity analyses, standardizing contracts generates larger savings with less variance than savings from single-payer strategies. CONCLUSION: Although moving toward a single-payer system will reduce BIR costs, certain reforms to payer-provider contracts could generate at least as many administrative cost savings without radically reforming the entire health system. BIR costs can be meaningfully reduced without abandoning a multi-payer system.

The Hispanic paradox in the prevalence of obesity at the county-level.

OBJECTIVE: The percentage of Hispanics in a county has a negative association with prevalence of obesity. Because Hispanic individuals are unevenly distributed in the United States, this study examined whether this protective association persists when stratifying counties into quartiles based on the size of the Hispanic population and after adjusting for county-level demographic, socioeconomic, healthcare, and environmental factors. METHODS: Data were extracted from the 2018 Robert Wood Johnson Foundation County Health Rankings. Counties were categorized into quartiles based on their percentage of Hispanics, 0%-5% (n = 1794), 5%-20% (n = 962), 20%-50% (n = 283), and >50% (n = 99). For each quartile, univariate and multivariate regression models were used to evaluate the association between prevalence of obesity and demographic, socioeconomic, healthcare, and environmental factors. RESULTS: Counties with the top quartile of Hispanic individuals had the lowest prevalence of obesity compared to counties at the bottom quartile (28.4 ± 3.6% vs. 32.7 ± 4.0%). There was a negative association between county-level percentage of Hispanics and prevalence of obesity in unadjusted analyses that persisted after adjusting for all county-level factors. CONCLUSIONS: Counties with a higher percentage of Hispanics have lower levels of obesity, even after controlling for demographic, socioeconomic, healthcare, and environmental factors. More research is needed to elucidate why having more Hispanics in a county may be protective against county-level obesity.

County-Level Factors Associated With Cardiovascular Mortality by Race/Ethnicity.

Background Persistent racial/ethnic disparities in cardiovascular disease (CVD) mortality are partially explained by healthcare access and socioeconomic, demographic, and behavioral factors. Little is known about the association between race/ethnicity-specific CVD mortality and county-level factors. Methods and Results Using 2017 county-level data, we studied the association between race/ethnicity-specific CVD age-adjusted mortality rate (AAMR) and county-level factors (demographics, census region, socioeconomics, CVD risk factors, and healthcare access). Univariate and multivariable linear regressions were used to estimate the association between these factors; R2 values were used to assess the factors that accounted for the greatest variation in CVD AAMR by race/ethnicity (non-Hispanic White, non-Hispanic Black, and Hispanic/Latinx individuals). There were 659 740 CVD deaths among non-Hispanic White individuals in 2698 counties; 100 475 deaths among non-Hispanic Black individuals in 717 counties; and 49 493 deaths among Hispanic/Latinx individuals across 267 counties. Non-Hispanic Black individuals had the highest mean CVD AAMR (320.04 deaths per 100 000 individuals), whereas Hispanic/Latinx individuals had the lowest (168.42 deaths per 100 000 individuals). The highest CVD AAMRs across all racial/ethnic groups were observed in the South. In unadjusted analyses, the greatest variation (R2) in CVD AAMR was explained by physical inactivity for non-Hispanic White individuals (32.3%), median household income for non-Hispanic Black individuals (24.7%), and population size for Hispanic/Latinx individuals (28.4%). In multivariable regressions using county-level factor categories, the greatest variation in CVD AAMR was explained by CVD risk factors for non-Hispanic White individuals (35.3%), socioeconomic factors for non-Hispanic Black (25.8%), and demographic factors for Hispanic/Latinx individuals (34.9%). Conclusions The associations between race/ethnicity-specific age-adjusted CVD mortality and county-level factors differ significantly. Interventions to reduce disparities may benefit from being designed accordingly.

The design and evaluation of a novel algorithm for automated preference card optimization.

BACKGROUND: Inaccurate surgical preference cards (supply lists) are associated with higher direct costs, waste, and delays. Numerous preference card improvement projects have relied on institution-specific, manual approaches of limited reproducibility. We developed and tested an algorithm to facilitate the first automated, informatics-based, fully reproducible approach. METHODS: The algorithm cross-references the supplies used in each procedure and listed on each preference card and uses a time-series regression to estimate the likelihood that each quantity listed on the preference card is inaccurate. Algorithm performance was evaluated by measuring changes in direct costs between preference cards revised with the algorithm and preference cards that were not revised or revised without use of the algorithm. Results were evaluated with a difference-in-differences (DID) multivariate fixed-effects model of costs during an 8-month pre-intervention and a 15-month post-intervention period. RESULTS: The accuracies of the quantities of 469 155 surgeon-procedure-specific items were estimated. Nurses used these estimates to revise 309 preference cards across eight surgical services corresponding to, respectively, 1777 and 3106 procedures in the pre- and post-intervention periods. The average direct cost of supplies per case decreased by 8.38% ($352, SD $6622) for the intervention group and increased by 13.21% ($405, SD $14 706) for the control group (P < .001). The DID analysis showed significant cost reductions only in the intervention group during the intervention period (P < .001). CONCLUSION: The optimization of preference cards with a variety of institution-specific, manually intensive approaches has led to cost savings. The automated algorithm presented here produced similar results that may be more readily reproducible.

Uninterrupted continuous glucose monitoring access is associated with a decrease in HbA1c in youth with type 1 diabetes and public insurance.

OBJECTIVE: Continuous glucose monitor (CGM) use is associated with improved glucose control. We describe the effect of continued and interrupted CGM use on hemoglobin A1c (HbA1c) in youth with public insurance. METHODS: We reviewed 956 visits from 264 youth with type 1 diabetes (T1D) and public insurance. Demographic data, HbA1c and two-week CGM data were collected. Youth were classified as never user, consistent user, insurance discontinuer, and self-discontinuer. Visits were categorized as never-user visit, visit before CGM start, visit after CGM start, visit with continued CGM use, visit with initial loss of CGM, visit with continued loss of CGM, and visit where CGM is regained after loss. Multivariate regression adjusting for age, sex, race, diabetes duration, initial HbA1c, and body mass index were used to calculate adjusted mean and delta HbA1c. RESULTS: Adjusted mean HbA1c was lowest for the consistent user group (HbA1c 8.6%;[95%CI 7.9,9.3]). Delta HbA1c (calculated from visit before CGM start) was lower for visit after CGM start (-0.39%;[95%CI -0.78,-0.02]) and visit with continued CGM use (-0.29%;[95%CI -0.61,0.02]), whereas it was higher for visit with initial loss of CGM (0.40%;[95%CI -0.06,0.86]), visit with continued loss of CGM (0.46%;[95%CI 0.06,0.85]), and visit where CGM is regained after loss (0.57%;[95%CI 0.06,1.10]). CONCLUSIONS: Youth with public insurance using CGM have improved HbA1c, but only when CGM use is uninterrupted. Interruptions in use, primarily due to gaps in insurance coverage of CGM, were associated with increased HbA1c. These data support both initial and ongoing coverage of CGM for youth with T1D and public insurance.

Identification of Factors Associated With Variation in US County-Level Obesity Prevalence Rates Using Epidemiologic vs Machine Learning Models.

Importance: Obesity is a leading cause of high health care expenditures, disability, and premature mortality. Previous studies have documented geographic disparities in obesity prevalence. Objective: To identify county-level factors associated with obesity using traditional epidemiologic and machine learning methods. Design, Setting, and Participants: Cross-sectional study using linear regression models and machine learning models to evaluate the associations between county-level obesity and county-level demographic, socioeconomic, health care, and environmental factors from summarized statistical data extracted from the 2018 Robert Wood Johnson Foundation County Health Rankings and merged with US Census data from each of 3138 US counties. The explanatory power of the linear multivariate regression and the top performing machine learning model were compared using mean R2 measured in 30-fold cross validation. Exposures: County-level demographic factors (population; rural status; census region; and race/ethnicity, sex, and age composition), socioeconomic factors (median income, unemployment rate, and percentage of population with some college education), health care factors (rate of uninsured adults and primary care physicians), and environmental factors (access to healthy foods and access to exercise opportunities). Main Outcomes and Measures: County-level obesity prevalence in 2018, its association with each county-level factor, and the percentage of variation in county-level obesity prevalence explained by linear multivariate and gradient boosting machine regression measured with R2. Results: Among the 3138 counties studied, the mean (range) obesity prevalence was 31.5% (12.8%-47.8%). In multivariate regressions, demographic factors explained 44.9% of variation in obesity prevalence; socioeconomic factors, 33.0%; environmental factors, 15.5%; and health care factors, 9.1%. The county-level factors with the strongest association with obesity were census region, median household income, and percentage of population with some college education. R2 values of univariate regressions of obesity prevalence were 0.238 for census region, 0.218 for median household income, and 0.160 for percentage of population with some college education. Multivariate linear regression and gradient boosting machine regression (the best-performing machine learning model) of obesity prevalence using all county-level demographic, socioeconomic, health care, and environmental factors had R2 values of 0.58 and 0.66, respectively (P < .001). Conclusions and Relevance: Obesity prevalence varies significantly between counties. County-level demographic, socioeconomic, health care, and environmental factors explain the majority of variation in county-level obesity prevalence. Using machine learning models may explain significantly more of the variation in obesity prevalence..

Improving the efficiency of the operating room environment with an optimization and machine learning model.

The operating room is a major cost and revenue center for most hospitals. Thus, more effective operating room management and scheduling can provide significant benefits. In many hospitals, the post-anesthesia care unit (PACU), where patients recover after their surgical procedures, is a bottleneck. If the PACU reaches capacity, patients must wait in the operating room until the PACU has available space, leading to delays and possible cancellations for subsequent operating room procedures. We develop a generalizable optimization and machine learning approach to sequence operating room procedures to minimize delays caused by PACU unavailability. Specifically, we use machine learning to estimate the required PACU time for each type of surgical procedure, we develop and solve two integer programming models to schedule procedures in the operating rooms to minimize maximum PACU occupancy, and we use discrete event simulation to compare our optimized schedule to the existing schedule. Using data from Lucile Packard Children's Hospital Stanford, we show that the scheduling system can significantly reduce operating room delays caused by PACU congestion while still keeping operating room utilization high: simulation of the second half of 2016 shows that our model could have reduced total PACU holds by 76% without decreasing operating room utilization. We are currently working on implementing the scheduling system at the hospital.