The integrated vulnerability measure a novel financial distress index for US hospitals Authors: Julianna Bernardi, Paul Ursino, Kevin Gormley, Colleen Rye, Matt Boyas, Sam Steckley Published: 2026 ABSTRACT AbstractIntroductionHospital closures and other manifestations of financial distress threaten timely access to care across the United States. Widely used, publicly available distress indicators have important limitations when applied across diverse hospital types and operating environments.ResultsThe IVM showed strong predictive power and sorted hospitals into meaningful risk groups. In the validation sample, 51% of hospitals were classified as low risk, 30% as mid-low risk, 11% as mid-high risk, and 8% as high risk. Higher-risk hospitals were substantially more likely than lower-risk hospitals to experience financial distress within two years.ConclusionThe IVM provides a flexible, generalizable tool for early identification of vulnerable hospitals and for targeting technical assistance, oversight, or interventions to preserve local access to essential care.Keywords: hospital financial distress, hospital closures, Healthcare Cost Report Information System, health economics, financial risk Introduction Hospitals are critical community assets, providing around-the-clock emergency services, inpatient care, and local employment. When a hospital enters financial distress, clinical services may be reduced, patients must travel farther for care, and regional economies can suffer.1 Since 2010,2 over 150 hospitals have closed, and many urban and rural facilities run on consistently thin margins.3 Identifying financially vulnerable facilities early is essential to protect both community health and economic stability. Several publicly available indicators already seek to measure hospital financial health. The rural financial distress index (FDI) model from the Rural Health Research Program (RHRP) at the University of North Carolina Cecil G. Sheps Center for Health Services Research (“Sheps”)4 is calibrated specifically for rural institutions. Adaptations of the Altman Z-score5 and the Ohlson O-score6 that were originally developed for other sectors have been applied in healthcare.7,8 While these indicators contribute valuable insight, they target specific hospital types and generally assume complete underlying data. Those assumptions can limit coverage, especially for hospitals with incomplete data, which are often the organizations that may warrant attention by community leaders. Recent work has also focused on predicting hospital bankruptcy using longitudinal hospital data and comparing modern models to legacy scores.9 The integrated vulnerability measure (IVM) is intended to be complementary to this work by forecasting a composite distress outcome that is observable in publicly available data and uses an interpretable random forest model. To provide a more comprehensive approach, IVM extended the FDI model to all general acute-care hospitals that filed Medicare cost reports from 2017 to 2022, including critical access hospital (CAH) and excluding local, state, and federal facilities. The IVM is intended for use within this hospital population and is not validated for other facility types. Financial and nonfinancial predictors were combined to detect potential distress 2 years in advance. This approach leveraged publicly available Centers for Medicare & Medicaid Services (CMS) Healthcare Cost Report Information System (HCRIS) data10 and used a random forest model to accommodate missing data. The resulting measure is intended to complement existing measures and offer policymakers, researchers, and community leaders an additional tool for proactive intervention. Methods Data sources Healthcare Cost Report Information System data,10 annualized to align with calendar years, was the primary data source, which covers nearly all Medicare-certified hospitals (approximately 98% of community hospitals11). Additional hospital and community characteristics were obtained from the US Census Bureau American Community Survey and CMS public use files. Defining hospital financial distress Three indicators were used to define hospital distress, in alignment with the FDI model4: (1) closure; (2) cash flow margin less than −5%, indicating operating deficiencies and the inability to generate sufficient cash from operations; and (3) negative equity (book value): where liabilities exceed assets, a sign of long-term solvency issues. A single financial distress measure was defined as the occurrence of any indicator 2 years after the observation year. Observations were treated as missing on the outcome only when all 3 distress indicators were missing. Closure data was compiled from the Sheps rural hospital closure list12 and Becker's Hospital Review,13 matched to HCRIS records. Sensitivity analyses were conducted to evaluate the distress outcome construction validity and elaborated on in the Discussion section. Data processing and measures Data submission in HCRIS may include incomplete, inconsistent, or erroneous data fields; these rows are often removed, which limit the set of usable records. The IVM applied additional data processing steps to retain observations while reducing the influence of extreme values. Missing values were not broadly imputed, and missingness was treated as informative for prediction. Forty-five financial and 40 nonfinancial variables were assembled, spanning financial performance and solvency, community characteristics, payer mix, and operational indicators. Lagged versions were created for a subset of these, yielding over 200 candidate features. The selection was then reduced to a parsimonious set that preserved predictive performance and interpretability. The resulting 24 explanatory variables are listed in Table 1. Additional detail on data processing, variable construction, and sensitivity analysis is provided in the Appendix. (To access the Appendix, click on the Details tab of the article online.) Table 1Distributions of model predictors for train and validation sets combined (2017-2022). Predictors General formula Mean (standard deviation) or % with outcome % missing Rural status (binary)FORHP rural/nonrural 38.43% 0.06% Negative net income (binary)Net income < 0 27.17% 0.01% Medicaid expansion status (binary)State has adopted Medicaid expansion by year t 64.57% 0.00% Critical hospital status (binary)CAH/not CAH 22.86% 0.00% For-profit status (binary)For-profit/not for-profit 22.66% 0.00% Change of ownership (CHOW) event (binary)Whether hospital participated in a CHOW transaction (as buyer or seller) in year t 1.14% 0.00% CMS quality scores: 30-day heart failure mortality rateEstimate of deaths within 30 days of the start of a hospital admission from heart failure 0.88 (0.02) 33.09% CMS quality scores: 30-day hospital-wide readmission rateEstimate of unplanned readmission to any acute-care hospital within 30 days of discharge from a hospitalization for any cause 15.15 (0.81) 4.75% Total margin (1-year lag)total margin, 1 year prior 0.06 (0.1) 2.52% Operating margin (1-year lag)(Net patient revenue − total operating expense)/net patient revenue (1 year prior) −0.01 (0.13) 2.13% Modified Altman Z X3: efficiency ratioEarnings before interest and taxes (EBITa)/total assets 0.07 (0.15) 1.19% Total marginNet income/(net patient revenue + total other income) 0.06 (0.1) 0.98% Return on equity (book value)Net income/equity (book value) 0.13 (0.3) 0.98% Equity (book value) ratioEquity (book value)/total assets 0.59 (0.51) 0.92% Medicare payer mixMedicare inpatient and outpatient charges/combined outpatient and inpatient total charges 0.24 (0.08) 0.65% Operating margin(Net patient revenue − total operating expense)/net patient revenue −0.02 (0.14) 0.64% Medicaid revenue shareNet revenue from Medicaid/net patient revenue 0.1 (0.07) 0.63% Occupancy rateTotal inpatient days/total bed days available 0.49 (0.21) 0.05% Service area population percentage uninsuredPercent population in hospital service area uninsured 9.09 (2.96) 0.04% Service area population percentage in povertyPercent population in hospital service area below poverty level 15 (3.88) 0.04% Service area population percentage femalePercent population in hospital service area female 50.78 (0.74) 0.04% Service area population percentage aged 65 and overPercent population in hospital service area aged 65 and over 16.5 (2.42) 0.04% Medicaid payer mixMedicaid inpatient and outpatient charges/combined outpatient and inpatient total charges 0.16 (0.08) 0.00% Number of hospitals within 50 milesNumber of neighboring hospitals within a 50-mile distance threshold, excluding the reference hospital 23.59 (23.66) 0.00% SOURCE Resulting analysis of data from Centers for Medicare & Medicaid Services (CMS) Healthcare Cost Report Information System (HCRIS), Clinical Outcomes Domain Scores, Hospital Readmissions Reduction Program (HRRP) Files, Change of Ownership Files, Medicare Advantage Penetration Files, and Hospital Service Area Files; and Federal Office of Rural Health Policy (FORHP) Data Files, United States Census Bureau American Community Survey 5-Year Data (2009-2023), and KFF Medicaid Expansion Status data. NOTES Definitions and sources are given in the Appendix. aEarnings before interest and taxes (EBIT) is not an available field in the HCRIS data, and taxes and interest are not reported consistently. Thus, EBIT was calculated as follows: [(total patient revenue − contractual allowances and discounts on patient accounts) − total operating expenses + total other income − total other expenses]. This follows Enumah and Chang (2021) but takes into account missing data to refine the formula. Distributions of model predictors for train and validation sets combined (2017-2022). SOURCE Resulting analysis of data from Centers for Medicare & Medicaid Services (CMS) Healthcare Cost Report Information System (HCRIS), Clinical Outcomes Domain Scores, Hospital Readmissions Reduction Program (HRRP) Files, Change of Ownership Files, Medicare Advantage Penetration Files, and Hospital Service Area Files; and Federal Office of Rural Health Policy (FORHP) Data Files, United States Census Bureau American Community Survey 5-Year Data (2009-2023), and KFF Medicaid Expansion Status data. NOTES Definitions and sources are given in the Appendix. aEarnings before interest and taxes (EBIT) is not an available field in the HCRIS data, and taxes and interest are not reported consistently. Thus, EBIT was calculated as follows: [(total patient revenue − contractual allowances and discounts on patient accounts) − total operating expenses + total other income − total other expenses]. This follows Enumah and Chang (2021) but takes into account missing data to refine the formula. Model framework A random forest model14 was used to estimate the probability that a hospital enters financial distress within 2 years. The study used a (unbalanced) longitudinal panel of hospital year observations identified at the CMS Certification Number (CCN) level, so hospitals contributed multiple observations over time. Thus, since hospital years are clustered within hospitals, they are not statistically independent. The model predicts distress in year t + 2 using information observed in year t (and selected lagged measures). The goal of the study is risk prediction rather than statistical inference, so we do not interpret model parameters as independent effect estimates. The model was trained on observations from the years 2017 to 2020 and validated on data from the years 2021 to 2022. Random forest was chosen because it can capture nonlinear relationships, provides interpretable feature importance, and can accommodate missing data without requiring direct imputation. Recent studies have also demonstrated random forest as a strong performer in healthcare-specific distress prediction compared to other machine learning approaches.15,16 Performance was assessed using 5-fold cross-validated root mean squared error (RMSE), area under the receiver operating characteristic curve (AUC), area under the precision–recall curve (AUPRC), and threshold-based precision, recall, and balanced accuracy. Hospital year observations were also grouped into 4 risk categories based on predicted probabilities: low risk, mid-low risk, mid-high risk, and high risk. Each category reflects a predicted probability less than or equal to 10%, greater than 10% and less than or equal to 30%, greater than 30% and less than or equal to 60%, and greater than 60%, respectively. (These cutoffs are heuristic, and future work should determine optimal thresholds.) Results Descriptive statistics In preparing data for analysis, 90 observations with 30 or more missing HCRIS variables were removed. A total of 2196 (10.6%) observations were excluded due to missing all 3 distress outcome measure data. (Seventy-one percent of those observations came from incomplete or missing 2024 data [used to determine 2022 outcomes for the validation set]. Since 2024 data falls under the remediation period for HCRIS reports, for which hospitals have an 18-month window to submit a cost report, this higher rate of missing data is expected. See the Appendix for more details.) One hospital was excluded because it is not operational despite active yearly cost reports.17 The final train and validation sets included 13 367 and 5182 observation-years, respectively. Most hospitals were small (0-80 beds) to medium-sized (81-299 beds) nonprofits (35% and 27%, respectively), with rural hospitals predominating among smaller facilities. From 10 years of hospital data, the median coverage per hospital was 9 years, (Hospitals with fewer years of data available either closed, switched type [e.g., switched to CAH designation], were acquired and changed CCN as a result, or became a new hospital after 2015.) with the 25th and 75th percentiles both at 9 years. A total of 19.2% of observations in the train and validation sets experienced financial distress. The hospitals are a superset of rural hospitals captured in FDI, increasing the predictive set by about 25%. Summary statistics are provided in Table 1 and detailed in the Appendix. Modeling results The model had average cross-validated RMSEs of 0.30 (train) and 0.31 (validation), balanced accuracy of 0.77, AUC 0.88, and AUPRC 0.73, representing strong predictive power.18,19 At a low probability threshold (0.1), the model accurately identifies 89% of hospitals experiencing financial distress in 2 years (recall) but has a precision of 0.37. At a higher threshold (0.5), recall is 0.43 and precision is 0.85. See the Appendix for a comprehensive list of performance metrics. Figure 1, displayed similarly to FDI,4 plots observed vs predicted distress rates. Each dot represents a prediction range, and the dashed diagonal marks perfect calibration. As shown, observed distress rates increase monotonically with predicted risk, and predicted and observed rates are aligned across most bins. Observed percentage vs predicted probability of financial distress for validation set (2021-2022). Source/Notes: SOURCE Authors’ analysis of random forest modeling output. NOTES Percentages are removed from each probability range subsample; sources and definitions for modeling input are outlined in Table 1. Distress index groups Each observation in the validation set was placed into 1 of the 4 risk categories: low (51%), mid-low (30%), mid-high (11%), and high risk (8%). Figure 2 reports the observed financial distress rate within each risk category in the validation set. To assess how these risk categories relate to the individual components of the composite distress outcome, Figure 3 presents the same stratification disaggregated by distress type, like that displayed for FDI.4 The differences in proportions across risk categories are larger for hospitals with negative equity (book value) than for the other indicators (especially closed hospitals). Implications are discussed in the Policy relevance and decision use section. Observed rates of financial distress by risk category for validation set (2021-2022). Source/Notes: SOURCE Authors’ analysis of random forest modeling output. NOTES Percentages are removed from each risk category subsample; sources and definitions for modeling input are outlined in Table 1. Observed rates of financial distress by risk category and distress indicator for validation set (2021-2022). Source/Notes: SOURCE Authors’ analysis of random forest modeling output. NOTES Percentages are removed from each risk category and distress indicator subsample; percentages will not add to 100 because a hospital can experience multiple types of distress at once; sources and definitions for modeling input are outlined in Table 1. Feature interpretability To improve interpretability, Shapley Additive Explanations (SHAP) values were calculated for a random sample of 2000 from the validation set. Results show that IVM predictions are driven primarily by 3 financial measures: operating margin, total margin, and equity (book value) ratio. Higher values of these measures generally reduce predicted distress risk, while weaker margins and especially negative equity increase risk. Other factors (payer mix, competition, and community characteristics) make smaller secondary adjustments. See the Appendix for further discussion. Discussion Policy relevance and decision use The IVM was designed to support early identification of hospitals at elevated risk of experiencing the study's financial distress outcome within 2 years. It is intended to provide a single, consistent screening framework across a broad set of hospitals (acute-care and CAH, rural and urban) using public data that may be incomplete. Financial distress index remains appropriate when decision-makers need a specific model calibrated to rural hospitals and rural operating conditions. Modified Altman Z remains useful as a simple, transparent ratio-based screen, but its applicability can be constrained when required fields are missing and when decision-makers want to incorporate nonfinancial context. For some applications, such as prioritizing a limited set of hospitals for follow-up review, decision-makers may also consider comparing signals across multiple tools as an informal concordance check, recognizing that different models are optimized for different populations and definitions of distress. The model produces predicted probabilities, which were placed into 4 risk categories using heuristic probability cutoffs. As shown in Figures 2 and 3, hospitals in higher IVM risk tiers experienced substantially higher observed distress rates in the validation set, exemplifying that the categories meaningfully separate facilities that may warrant closer attention. The probability cutoffs were intentionally not formally optimized because the preferred threshold depends on the policy objective and the relative costs of false positives vs false negatives. The clear precision–recall tradeoff across probability thresholds, as shown in Figure S8, demonstrates that lower thresholds (eg, 0.15) identify a larger share of hospitals that will experience distress (recall 0.82) but generate more false positives (precision 0.45), which may be appropriate for a state health department using the IVM for broad surveillance or monitoring to prompt data validation and situational awareness activities, such as routine reviews. Higher thresholds (eg, 0.5) yield a smaller, more targeted set of hospitals (precision 0.85) with fewer false positives but miss more distressed hospitals (recall 0.43), which may be more appropriate for regional planners or policymakers when prioritizing a smaller subset of hospitals for resource-intensive actions, such as technical assistance, contingency planning discussions, or enhanced oversight engagement. These examples are intended to illustrate how thresholds can be selected in relation to false-positive vs false-negative tradeoffs, rather than to prescribe a single decision rule. Figure 1 provides calibration context for model applicability. Distress rates rise with higher predicted probability ranges and track closely with diagonal line overall, while the model tends to underpredict distress at the upper end (Figure S12). This reinforces interpreting high predicted values as a strong screening signal while recognizing that predicted probabilities may be conservative at the upper end. The strongest predictors of the model based on SHAP values—operating margin, total margin, and equity (book value) ratio— align with core dimensions of hospital financial condition discussed in the health economics literature, particularly operating profitability and balance-sheet solvency/leverage. This result provides face validity as the algorithm detects metrics that financial stakeholders monitor when assessing hospital distress risk. Literature comparison The results suggest that the IVM can complement existing approaches for early detection of hospital financial distress and performs favorably relative to selected publicly available measures evaluated in this study. The model AUC (0.88) closely matches that of FDI.4 Additionally, the model demonstrates notable improvements over the modified Altman Z in identifying distressed hospitals. At a probability cutoff of 0.3 in the validation set, the model identified about 22% more distressed facilities than the modified Altman Z and showed higher balanced accuracy, precision, and recall. Additionally, Altman Z cannot be calculated for 78 validation set hospitals because of missing component variables, whereas the approach in this study covers every facility. While the modified Altman Z differentiates among safe, gray, and distressed categories, (10%, 24%, and 59% of observation-years in these categories were truly distressed for the validation set, respectively.) the adopted approach further separates a high-risk group that, in the validation set, had the highest observed distress rate. This may help decision-makers distinguish hospitals warranting more immediate follow-up from those better suited to ongoing monitoring. Real-world examples Recent hospital failures were compared to IVM predicted probabilities. Martin General Hospital in North Carolina had predicted distress probabilities of 0.77 for 2021 and 0.89 for 2022; the hospital filed for bankruptcy and closed in late 2023,20 confirming the predictive capacity of the model more than a year in advance. Steward Carney Hospital in Massachusetts, with probabilities of 0.76 for 2021 and 0.81 for 2022, went bankrupt and closed in 2024.21 These examples illustrate that the IVM assigned high predicted risk to hospitals that subsequently experienced severe financial distress events, including bankruptcy filings and closure. Limitations These results are limited to general short-term acute-care and CAH and may not generalize to other facility types. Healthcare Cost Report Information System data can be delayed and incomplete.11 Data is self-reported and may reflect system-level rather than hospital-level finances, (Although this has not been reported officially and HCRIS cost reports are required at the hospital level, the analyses in this study suggest that hospitals report at the system level for some variables. This is especially acute for total assets and total liabilities. For example, both providers under the Arnot Health system in New York—IRA Davenport Memorial Hospital (provider CCN: 330144) and Arnot-Ogden Medical Center (provider CCN: 330090)—report identical total assets on their cost reports between 2016 and 2023. Further, IRA Davenport Memorial Hospital is a 15-bed facility, whereas Arnot-Ogden Medical Center has over 190 beds, indicating that the duplication is likely due to aggregate statistics and not facility-level metrics.), which can distort hospital-specific predictors and introduce generalizability issues. In hospital systems, financial performance may be obscured if struggling member hospitals report system-level metrics, or conversely may appear more distressed if poor system-level metrics are reported for specific member facilities. As a result, the IVM should be interpreted as a risk signal derived from HCRIS cost report data, and predictions for system-affiliated hospitals may require additional operational context. Future work should more rigorously detect and adjust for system-level reporting artifacts and test whether model calibration differs by system affiliation. Because the relative costs of false positives vs false negatives vary by policy objective, probability thresholds and risk tiers should be interpreted as illustrative communication aids rather than prescriptive triggers for action. The financial distress definition used for this study relies on balance-sheet proxies, which may lag or only partially capture real-time hardship, as HCRIS does not directly report events such as bankruptcies or service-line closures. The IVM discriminates negative equity more strongly than closures and negative cash flow margin less than −5%, as shown in Figure 3. This is expected given that closures are very rare and cash flow margins are much noisier (Figure S13), while negative equity is more prevalent and more consistently measured in cost reports. To assess whether this result was due to the model being primarily driven by balance-sheet solvency information, the model was refit excluding the equity (book value) ratio predictor. The AUC declined from 0.877 to 0.863 and balanced accuracy from 0.774 to 0.758. Feature importance shifted toward operating performance measures (eg, efficiency and margins). Closure remained difficult to discriminate due to its low prevalence; however, the high-risk category continued to identify a small subset of hospitals with elevated closure rates (from 3.8% to 5.5%). Together, these findings suggest the IVM functions primarily as an early-warning measure of severe financial vulnerability as defined, including solvency risk. More details on this analysis are given in the Appendix. To better characterize the model's poorer discrimination of hospitals that experienced negative cash flow margin less than −5% in 2 years, changes in cash flow margin between the observation year and 2 years later were compared across distress types. A more negative shift and more extreme declines were observed for the negative cash flow margin less than −5% only subgroup than for the no-distress subgroup, for which changes were more tightly centered around 0. This pattern suggests that cash flow margin distress was more strongly characterized by subsequent deterioration over the 2-year horizon than by a clearly distinct profile in the available cost report signals, which may have limited advance identification of these cases. The change in cash flow margin comparison is shown in Figure S16. Certain data elements may introduce random variation. Rural status is determined at the county level and may not reflect historical changes. Ownership changes are only logged for currently Medicare-approved hospitals. (As of the last quarter of 2024, the date for which data used in this study was downloaded from CMS.) (CMS does include historical change of ownership files going back to 2022, but since this study period goes back to 2017 and it is rare for hospitals to lose enrollment or close, data processing was simplified to only include the most recent file.) Some closed hospitals continue to file minimal cost reports; despite removing clear cases, undetected cases may persist. Due to the COVID-19 pandemic, the year 2020 was an anomaly, but data from that year was retained to ensure model robustness across both typical and shock periods. However, including 2020 may affect the generalizability of results to nonpandemic years and introduce patterns that do not reflect normal operational conditions. See the Appendix for discussion on the inclusion of the 2020 data. The analytic methods used for this study have inherent caveats and assumptions: Treating missing values as informative may reduce predictive power and should be interpreted as a predictive signal, not as a causal driver of distress; distress groups are based on heuristic thresholds; and feature importance scores should not be interpreted as causal effects. Future work Alternative data sources, such as quarterly filings, state-level notices of service-line closures, or proprietary sources, could provide more timely data; could fill gaps in staffing, service mix, and utilization that are not captured in cost reports alone; and could materially improve calibration. Class imbalance, the Model Framework, and outlier handling are important considerations for model training and performance. Future work could include synthetic sampling methods, such as SMOTE22 or ADASYN,23 paired with class-weighted loss functions to account for class imbalances. Gradient-boosted trees, deep learning models, or stacked ensembles could capture higher-order interactions; their calibration should be verified within strata (eg, rural vs urban). Winsorization of extreme values could be tuned or replaced with adaptive weighting or quantile random forest to preserve informative tail events. Periodic recalibration and further research will be essential to understanding trends as major changes to the health system, such as payment policy, labor costs, and case-mix patterns, could evolve, limiting the generalizability of the results. Conclusion Health facilities will remain under financial pressures that can directly impact the communities served. Thus, it is important to continue to understand the financial health of hospitals. The distress index compares favorably with selected hospital distress measures discussed here and extends a common screening framework across general acute-care hospitals and CAHs using publicly available data and a flexible modeling approach. In this study, the IVM showed discrimination similar to the published FDI results, outperformed the modified Altman Z on several validation metrics at the chosen cutoff, generated predictions for all hospitals in the analytic validation sample despite missing predictors, and offered a 4-tier risk scale that can be aligned with different decision contexts. Because the model relies on publicly available data and standard open-source software, it can be replicated, audited, and updated as market conditions evolve. While additional validation is warranted, the adopted approach offers researchers and policymakers a practical complement to existing, proprietary scoring systems and underscores the broader need for timelier, more granular financial reporting. Supplementary Material Acknowledgments The views, opinions and/or findings contained in this report are those of The MITRE Corporation and should not be construed as an official government position, policy, or decision, unless designated by other documentation. Approved for Public Release; Distribution Unlimited. Public Release Case Number 25-2653. The team is grateful for MITRE leadership support, who encouraged and continuously supported the work. The authors would like to thank Dr. Carla Bezold and Dr. Matthew McQueen for their invaluable expertise, encouragement, and support. Additionally, the team is grateful to the MITRE HealthToolkit team for streamlined data processing and integration. Contributor Information Julianna Bernardi, The MITRE Corporation, Bedford, MA 01730, United States. Paul Ursino, PwC, Boston, MA 02210, United States. Kevin Gormley, ECS, Fairfax, VA 22031, United States. Colleen Rye, Rye & Associates, Atlanta, GA, United States. Matt Boyas, Linksbridge SPC, Seattle, WA 98109, United States. Sam Steckley, The MITRE Corporation, Bedford, MA 01730, United States. Supplementary material Supplementary material is available at Health Affairs Scholar online. Funding This work was funded by The MITRE Corporation through the MITRE Independent Research and Development Program (Project IRD102.25.R4.VQA). Conflicts of interest Please see ICMJE form(s) for author conflicts of interest. These have been provided as supplementary materials. None of the authors have any conflicts of interest with respect to this study and paper. Notes Associated Data Supplementary Materials