Clinical risk prediction using language models: benefits and considerations.

OBJECTIVE: The use of electronic health records (EHRs) for clinical risk prediction is on the rise. However, in many practical settings, the limited availability of task-specific EHR data can restrict the application of standard machine learning pipelines. In this study, we investigate the potential of leveraging language models (LMs) as a means to incorporate supplementary domain knowledge for improving the performance of various EHR-based risk prediction tasks. METHODS: We propose two novel LM-based methods, namely "LLaMA2-EHR" and "Sent-e-Med." Our focus is on utilizing the textual descriptions within structured EHRs to make risk predictions about future diagnoses. We conduct a comprehensive comparison with previous approaches across various data types and sizes. RESULTS: Experiments across 6 different methods and 3 separate risk prediction tasks reveal that employing LMs to represent structured EHRs, such as diagnostic histories, results in significant performance improvements when evaluated using standard metrics such as area under the receiver operating characteristic (ROC) curve and precision-recall (PR) curve. Additionally, they offer benefits such as few-shot learning, the ability to handle previously unseen medical concepts, and adaptability to various medical vocabularies. However, it is noteworthy that outcomes may exhibit sensitivity to a specific prompt. CONCLUSION: LMs encompass extensive embedded knowledge, making them valuable for the analysis of EHRs in the context of risk prediction. Nevertheless, it is important to exercise caution in their application, as ongoing safety concerns related to LMs persist and require continuous consideration.

Assessing the Impact of the Center for Medicare and Medicaid Services Policy Guidance on Part D Prescriptions Among Hospice Patients.

Hospice care facilities are required to provide prescription drugs related to a hospice patient's terminal illness. From October 2010 to present, the Center for Medicare and Medicaid Services (CMS) has issued a series of communications regarding Medicare paying for hospice patients' prescription drugs under Part D that should be covered under the hospice Medicare Part A benefit. On April 4, 2011, CMS issued specific policy guidance to providers aimed at preventing inappropriate billing. While CMS has documented Part D prescription decreases in hospice patients, no research exists that connects these decreases and the policy guidance. This study aims to evaluate the effect of the April 4, 2011, policy guidance on hospice patients' Part D prescriptions. This study employed generalized estimating equations to assess (1) total monthly average prescriptions of all medications and (2) four categories of commonly prescribed hospice medications in pre-and-post policy guidance. This research used the Medicare claims of 113,260 Part D-enrolled Medicare male patients aged 66 and older between April 2009 and March 2013, including 110,547 non-hospice patients and 2713 hospice patients. Hospice patients' monthly average total Part D prescriptions decreased from 7.3 pre-policy guidance to 6.5 medications following the issuing of the guidance, while the four categories of hospice-specific medications decreased from .57 to .49. The findings of this study show that CMS's guidance issued to providers to prevent the inappropriate billing of hospice patients' prescriptions to the Part D benefit may lead to Part D prescription decreases as observed in this sample.

Mediation of an association between neighborhood socioeconomic environment and type 2 diabetes through the leisure-time physical activity environment in an analysis of three independent samples.

INTRODUCTION: Inequitable access to leisure-time physical activity (LTPA) resources may explain geographic disparities in type 2 diabetes (T2D). We evaluated whether the neighborhood socioeconomic environment (NSEE) affects T2D through the LTPA environment. RESEARCH DESIGN AND METHODS: We conducted analyses in three study samples: the national Veterans Administration Diabetes Risk (VADR) cohort comprising electronic health records (EHR) of 4.1 million T2D-free veterans, the national prospective cohort REasons for Geographic and Racial Differences in Stroke (REGARDS) (11 208 T2D free), and a case-control study of Geisinger EHR in Pennsylvania (15 888 T2D cases). New-onset T2D was defined using diagnoses, laboratory and medication data. We harmonized neighborhood-level variables, including exposure, confounders, and effect modifiers. We measured NSEE with a summary index of six census tract indicators. The LTPA environment was measured by physical activity (PA) facility (gyms and other commercial facilities) density within street network buffers and population-weighted distance to parks. We estimated natural direct and indirect effects for each mediator stratified by community type. RESULTS: The magnitudes of the indirect effects were generally small, and the direction of the indirect effects differed by community type and study sample. The most consistent findings were for mediation via PA facility density in rural communities, where we observed positive indirect effects (differences in T2D incidence rates (95% CI) comparing the highest versus lowest quartiles of NSEE, multiplied by 100) of 1.53 (0.25, 3.05) in REGARDS and 0.0066 (0.0038, 0.0099) in VADR. No mediation was evident in Geisinger. CONCLUSIONS: PA facility density and distance to parks did not substantially mediate the relation between NSEE and T2D. Our heterogeneous results suggest that approaches to reduce T2D through changes to the LTPA environment require local tailoring.

Predicting 6-month mortality of patients from their medical history: Comparison of multimorbidity index to Deyo-Charlson index.

While every disease could affect a patient's prognosis, published studies continue to use indices that include a selective list of diseases to predict prognosis, which may limit its accuracy. This paper compares 6-month mortality predicted by a multimorbidity index (MMI) that relies on all diagnoses to the Deyo version of the Charlson index (DCI), a popular index that utilizes a selective set of diagnoses. In this retrospective cohort study, we used data from the Veterans Administration Diabetes Risk national cohort that included 6,082,018 diabetes-free veterans receiving primary care from January 1, 2008 to December 31, 2016. For the MMI, 7805 diagnoses were assigned into 19 body systems, using the likelihood that the disease will increase risk of mortality. The DCI used 17 categories of diseases, classified by clinicians as severe diseases. In predicting 6-month mortality, the cross-validated area under the receiver operating curve for the MMI was 0.828 (95% confidence interval of 0.826-0.829) and for the DCI was 0.749 (95% confidence interval of 0.748-0.750). Using all available diagnoses (MMI) led to a large improvement in accuracy of predicting prognosis of patients than using a selected list of diagnosis (DCI).

Multi-omic molecular profiling guide's efficacious treatment selection in refractory metastatic breast cancer: a prospective phase II clinical trial.

This prospective phase II clinical trial (Side Out 2) explored the clinical benefits of treatment selection informed by multi-omic molecular profiling (MoMP) in refractory metastatic breast cancers (MBCs). Core needle biopsies were collected from 32 patients with MBC at trial enrollment. Patients had received an average of 3.94 previous lines of treatment in the metastatic setting before enrollment in this study. Samples underwent MoMP, including exome sequencing, RNA sequencing (RNA-Seq), immunohistochemistry, and quantitative protein pathway activation mapping by Reverse Phase Protein Microarray (RPPA). Clinical benefit was assessed using the previously published growth modulation index (GMI) under the hypothesis that MoMP-selected therapy would warrant further investigation for GMI ≥ 1.3 in ≥ 35% of the patients. Of the 32 patients enrolled, 29 received treatment based on their MoMP and 25 met the follow-up criteria established by the trial protocol. Molecular information was delivered to the tumor board in a median time frame of 14 days (11-22 days), and targetable alterations for commercially available agents were found in 23/25 patients (92%). Of the 25 patients, 14 (56%) reached GMI ≥ 1.3. A high level of DNA topoisomerase I (TOPO1) led to the selection of irinotecan-based treatments in 48% (12/25) of the patients. A pooled analysis suggested clinical benefit in patients with high TOPO1 expression receiving irinotecan-based regimens (GMI ≥ 1.3 in 66.7% of cases). These results confirmed previous observations that MoMP increases the frequency of identifiable actionable alterations (92% of patients). The MoMP proposed allows the identification of biomarkers that are frequently expressed in MBCs and the evaluation of their role as predictors of response to commercially available agents. Lastly, this study confirmed the role of MoMP for informing treatment selection in refractory MBC patients: more than half of the enrolled patients reached a GMI ≥ 1.3 even after multiple lines of previous therapies for metastatic disease.

Longitudinal Analysis of Neighborhood Food Environment and Diabetes Risk in the Veterans Administration Diabetes Risk Cohort.

Importance: Diabetes causes substantial morbidity and mortality among adults in the US, yet its incidence varies across the country, suggesting that neighborhood factors are associated with geographical disparities in diabetes. Objective: To examine the association between neighborhood food environment and risk of incident type 2 diabetes across different community types (high-density urban, low-density urban, suburban, and rural). Design, Setting, and Participants: This is a national cohort study of 4 100 650 US veterans without type 2 diabetes. Participants entered the cohort between 2008 and 2016 and were followed up through 2018. The median (IQR) duration of follow-up was 5.5 (2.6-9.8) person-years. Data were obtained from Veterans Affairs electronic health records. Incident type 2 diabetes was defined as 2 encounters with type 2 diabetes International Classification of Diseases, Ninth Revision or Tenth Revision codes, a prescription for diabetes medication other than metformin or acarbose alone, or 1 encounter with type 2 diabetes International Classification of Diseases Ninth Revision or Tenth Revision codes and 2 instances of elevated hemoglobin A1c (≥6.5%). Data analysis was performed from October 2020 to March 2021. Exposures: Five-year mean counts of fast-food restaurants and supermarkets relative to other food outlets at baseline were used to generate neighborhood food environment measures. The association between food environment and time to incident diabetes was examined using piecewise exponential models with 2-year interval of person-time and county-level random effects stratifying by community types. Results: The mean (SD) age of cohort participants was 59.4 (17.2) years. Most of the participants were non-Hispanic White (2 783 756 participants [76.3%]) and male (3 779 555 participants [92.2%]). The relative density of fast-food restaurants was positively associated with a modestly increased risk of type 2 diabetes in all community types. The adjusted hazard ratio (aHR) was 1.01 (95% CI, 1.00-1.02) in high-density urban communities, 1.01 (95% CI, 1.01-1.01) in low-density urban communities, 1.02 (95% CI, 1.01-1.03) in suburban communities, and 1.01 (95% CI, 1.01-1.02) in rural communities. The relative density of supermarkets was associated with lower type 2 diabetes risk only in suburban (aHR, 0.97; 95% CI, 0.96-0.99) and rural (aHR, 0.99; 95% CI, 0.98-0.99) communities. Conclusions and Relevance: These findings suggest that neighborhood food environment measures are associated with type 2 diabetes among US veterans in multiple community types and that food environments are potential avenues for action to address the burden of diabetes. Tailored interventions targeting the availability of supermarkets may be associated with reduced diabetes risk, particularly in suburban and rural communities, whereas restrictions on fast-food restaurants may help in all community types.

US veterans administration diabetes risk (VADR) national cohort: cohort profile.

PURPOSE: The veterans administration diabetes risk (VADR) cohort facilitates studies on temporal and geographic patterns of pre-diabetes and diabetes, as well as targeted studies of their predictors. The cohort provides an infrastructure for examination of novel individual and community-level risk factors for diabetes and their consequences among veterans. This cohort also establishes a baseline against which to assess the impact of national or regional strategies to prevent diabetes in veterans. PARTICIPANTS: The VADR cohort includes all 6 082 018 veterans in the USA enrolled in the veteran administration (VA) for primary care who were diabetes-free as of 1 January 2008 and who had at least two diabetes-free visits to a VA primary care service at least 30 days apart within any 5-year period since 1 January 2003, or veterans subsequently enrolled and were diabetes-free at cohort entry through 31 December 2016. Cohort subjects were followed from the date of cohort entry until censure defined as date of incident diabetes, loss to follow-up of 2 years, death or until 31 December 2018. FINDINGS TO DATE: The incidence rate of type 2 diabetes in this cohort of over 6 million veterans followed for a median of 5.5 years (over 35 million person-years (PY)) was 26 per 1000 PY. During the study period, 8.5% of the cohort were lost to follow-up and 17.7% died. Many demographic, comorbidity and other clinical variables were more prevalent among patients with incident diabetes. FUTURE PLANS: This cohort will be used to study community-level risk factors for diabetes, such as attributes of the food environment and neighbourhood socioeconomic status via geospatial linkage to residence address information.

Relative accuracy of social and medical determinants of suicide in electronic health records.

OBJECTIVE: This paper compares the accuracy of predicting suicide from Social Determinants of Health (SDoH) or history of illness. POPULATION STUDIED: 5 313 965 Veterans who at least had two primary care visits between 2008 and 2016. STUDY DESIGN: The dependent variable was suicide or intentional self-injury. The independent variables were 10 495 International Classification of Disease (ICD) Version 9 codes, age, and gender. The ICD codes included 40 V-codes used for measuring SDoH, such as family disruption, family history of substance abuse, lack of education, legal impediments, social isolation, unemployment, and homelessness. The sample was randomly divided into training (90 percent) and validation (10 percent) sets. Area under the receiver operating characteristic (AROC) was used to measure accuracy of predictions in the validation set. PRINCIPAL FINDINGS: Separate analyses were done for inpatient and outpatient codes; the results were similar. In the hospitalized group, the mean age was 67.2 years, and 92.1 percent were male. The mean number of medical diagnostic codes during the study period was 37; and 12.9 percent had at least one SDoH V-code. At least one episode of suicide or intentional self-injury occurred in 1.89 percent of cases. SDoH V-codes, on average, elevated the risk of suicide or intentional self-injury by 24-fold (ranging from 4- to 86-fold). An index of 40 SDoH codes predicted suicide or intentional self-injury with an AROC of 0.64. An index of 10 445 medical diagnoses, without SDoH V-codes, had AROC of 0.77. The combined SDoH and medical diagnoses codes also had AROC of 0.77. CONCLUSION: In predicting suicide or intentional self-harm, SDoH V-codes add negligible information beyond what is already available in medical diagnosis codes. IMPLICATIONS FOR PRACTICE: Policies that affect SDoH (eg, housing policies, resilience training) may not have an impact on suicide rates, if they do not change the underlying medical causes of SDoH.

Predictors of Clostridioides difficile recurrence across a national cohort of veterans in outpatient, acute, and long-term care settings.

PURPOSE: The greatest challenge in treating Clostridioides difficile infection (CDI) is disease recurrence, which occurs in about 20% of patients, usually within 30 days of treatment cessation. We sought to identify independent predictors of first recurrence among a national cohort of veterans with CDI. METHODS: We conducted a case-control study among acute and long-term care Veterans Affairs (VA) inpatients and outpatients with a first CDI episode (positive stool sample for C. difficile toxin[s] and receipt of at least 2 days of CDI treatment) between 2010 and 2014. Cases experienced first recurrence within 30 days from the end of treatment. Controls were those without first recurrence matched 4:1 to cases on year, facility, and severity. Multivariable conditional logistic regression was used to identify predictors of first recurrence. RESULTS: We identified 32 predictors of first recurrence among 974 cases and 3,896 matched controls. Significant predictors included medication use prior to (probiotics, fluoroquinolones, laxatives, third- or fourth-generation cephalosporins), during (first- or second-generation cephalosporins, penicillin/amoxicillin/ampicillin, third- and fourth-generation cephalosporins), and after CDI treatment (probiotics, any antibiotic, proton pump inhibitors [PPIs], and immunosuppressants). Other predictors included current biliary tract disease, malaise/fatigue, cellulitis/abscess, solid organ cancer, medical history of HIV, multiple myeloma, abdominal pain, and ulcerative colitis. CONCLUSION: In a large national cohort of outpatient and acute and long-term care inpatients, treatment with certain antibiotics, PPIs, immunosuppressants, and underlying disease were among the most important risk factors for first CDI recurrence.

Prior antibiotic exposure and risk of type 2 diabetes among Veterans.

BACKGROUND: Exposure to antibiotics may increase the risk of type 2 diabetes. Veterans are at increased risk for diabetes and for exposure to antibiotics. OBJECTIVE: To determine the impact of antibiotic exposure for risk of diabetes. DESIGN: Retrospective cohort study. PARTICIPANTS: Veterans at the New York Harbor Healthcare System enrolled in primary care, 2004-2014, with ≥2 glycosylated hemoglobin test results <6.5%. MAIN MEASURES: The primary exposure was any antimicrobial prescribed >6 months prior to the date of diabetes diagnosis, loss to follow-up, death, or the end of the study, measured as the number of courses of antimicrobial prescriptions filled and the mean daily dose (MDD). The primary outcome was incident diagnosis of diabetes through 2014, defined ≥2 ICD-9 codes for diabetes or ≥2 prescriptions of diabetes medications, other than metformin. Cox proportional hazards regression was used to model antimicrobial medications, demographic and anthropometric measures, and comorbid cardiovascular conditions to incident diabetes. Models incorporated time varying covariates of antimicrobial medication and MDD to analyze associations by antimicrobial class. KEY RESULTS: Among 14,361 Veterans, 9922 (69.1%) were prescribed any antimicrobial medication during the study period. 1413 (9.8%) individuals developed type 2 diabetes. Increased risk of diabetes was associated with >1 prescription (HR 1.13 [1.01-1.26]) compared to none. Time varying analysis of the total number of cumulative courses prescribed showed increased diabetes risk for cephalosporin (HR 1.17 [1.04-1.31]), macrolide (HR 1.08 [1.03-1.13]) and penicillin (HR 1.05 [1.02-1.07]). MDD showed increased risk per 100-unit (mg) increase in antibiotic exposure from (HR 1.05 [1.02-1.08]) for sulfonamide to (HR 1.70 [1.51-1.92]) for cephalosporin. CONCLUSION: Any and repeated exposure to certain antibiotics may increase diabetes risk among Veterans. Results from this study add to the growing evidence suggesting that antibiotic exposure increases risk for diabetes. Antibiotic stewardship may be enhanced by better understanding this risk, and may lower the incidence of diabetes in populations at risk.

Predictors of Mortality Among a National Cohort of Veterans With Recurrent Clostridium difficile Infection.

BACKGROUND: Though recurrent Clostridium difficile infection (CDI) is common and poses a major clinical concern, data are lacking regarding mortality among patients who survive their initial CDI and have subsequent recurrences. Risk factors for mortality in patients with recurrent CDI are largely unknown. METHODS: Veterans Affairs patients with a first CDI (stool sample with positive C. difficile toxin(s) and ≥2 days CDI treatment) were included (2010-2014). Subsequent recurrences were defined as additional CDI episodes ≥14 days after the stool test date and within 30 days of the end of treatment. A matched (1:4) case-control analysis was conducted using multivariable conditional logistic regression to identify predictors of all-cause mortality within 30 days of the first recurrence. RESULTS: Crude 30-day all-cause mortality rates were 10.6% for the initial CDI episode, 8.3% for the first recurrence, 4.2% for the second recurrence, and 5.9% for the third recurrence. Among 110 cases and 440 controls, 6 predictors of mortality were identified: use of proton pump inhibitors (PPIs; odds ratio [OR], 3.86; 95% confidence interval [CI], 2.14-6.96), any antibiotic (OR, 3.33; 95% CI, 1.79-6.17), respiratory failure (OR, 8.26; 95% CI, 1.71-39.92), congitive dysfunction (OR, 2.41; 95% CI, 1.02-5.72), nutrition deficiency (OR, 2.91; 95% CI, 1.37-6.21), and age (OR, 1.04; 95% CI, 1.01-1.07). CONCLUSIONS: In our national cohort of Veterans, crude mortality decreased by 44% from the initial episode to the third recurrence. Treatment with antibiotics, use of PPIs, and underlying comorbidities were important predictors of mortality in recurrent CDI. Our study assists health care providers in identifying patients at high risk of death after CDI recurrence.

Electronic Health Record-Based Screening for Substance Abuse.

Existing methods of screening for substance abuse (standardized questionnaires or clinician's simply asking) have proven difficult to initiate and maintain in primary care settings. This article reports on how predictive modeling can be used to screen for substance abuse using extant data in electronic health records (EHRs). We relied on data available through Veterans Affairs Informatics and Computing Infrastructure (VINCI) for the years 2006 through 2016. We focused on 4,681,809 veterans who had at least two primary care visits; 829,827 of whom had a hospitalization. Data included 699 million outpatient and 17 million inpatient records. The dependent variable was substance abuse as identified from 89 diagnostic codes using the Agency for Healthcare Quality and Research classification of diseases. In addition, we included the diagnostic codes used for identification of prescription abuse. The independent variables were 10,292 inpatient and 13,512 outpatient diagnoses, plus 71 dummy variables measuring age at different years between 20 and 90 years. A modified naive Bayes model was used to aggregate the risk across predictors. The accuracy of the predictions was examined using area under the receiver operating characteristic (AROC) curve in 20% of data, randomly set aside for the evaluation. Many physical/mental illnesses were associated with substance abuse. These associations supported findings reported in the literature regarding the impact of substance abuse on various diseases and vice versa. In randomly set-aside validation data, the model accurately predicted substance abuse for inpatient (AROC = 0.884), outpatient (AROC = 0.825), and combined inpatient and outpatient (AROC = 0.840) data. If one excludes information available after substance abuse is known, the cross-validated AROC remained high, 0.822 for inpatient and 0.817 for outpatient data. Data within EHRs can be used to detect existing or predict potential future substance abuse.

HbA1c, lipid profiles and risk of incident type 2 Diabetes in United States Veterans.

United States Veterans are at excess risk for type 2 diabetes, but population differentials in risk have not been characterized. We determined risk of type 2 diabetes in relation to prediabetes and dyslipidemic profiles in Veterans at the VA New York Harbor (VA NYHHS) during 2004-2014. Prediabetes was based on American Diabetes Association hemoglobin A1c (HbA1c) testing cut-points, one of several possible criteria used to define prediabetes. We evaluated transition to type 2 diabetes in 4,297 normoglycemic Veterans and 7,060 Veterans with prediabetes. Cox proportional hazards regression was used to relate HbA1c levels, lipid profiles, demographic, anthropometric and comorbid cardiovascular factors to incident diabetes (Hazard Ratio [HR] and 95% confidence intervals). Compared to normoglycemic Veterans (HbA1c: 5.0-5.6%; 31-38 mmol/mol), risks for diabetes were >2-fold in the moderate prediabetes risk group (HbA1c: 5.7-5.9%; 39-41 mmol/mol) (HR 2.37 [1.98-2.85]) and >5-fold in the high risk prediabetes group (HbA1c: 6.0-6.4%; 42-46 mmol/mol) (HR 5.59 [4.75-6.58]). Risks for diabetes were increased with elevated VLDL (≥40mg/dl; HR 1.31 [1.09-1.58]) and TG/HDL (≥1.5mg/dl; HR 1.34 [1.12-1.59]), and decreased with elevated HDL (≥35mg/dl; HR 0.80 [0.67-0.96]). Transition to diabetes in Veterans was related in age-stratified risk score analyses to HbA1c, VLDL, HDL and TG/HDL, BMI, hypertension and race, with 5-year risk differentials of 62% for the lowest (5-year risk, 13.5%) vs. the highest quartile (5-year risk, 21.9%) of the risk score. This investigation identified substantial differentials in risk of diabetes in Veterans, based on a readily-derived risk score suitable for risk stratification for type 2 diabetes prevention.

Predicting activities of daily living for cancer patients using an ontology-guided machine learning methodology.

BACKGROUND: Bio-ontologies are becoming increasingly important in knowledge representation and in the machine learning (ML) fields. This paper presents a ML approach that incorporates bio-ontologies and its application to the SEER-MHOS dataset to discover patterns of patient characteristics that impact the ability to perform activities of daily living (ADLs). Bio-ontologies are used to provide computable knowledge for ML methods to "understand" biomedical data. RESULTS: This retrospective study included 723 cancer patients from the SEER-MHOS dataset. Two ML methods were applied to create predictive models for ADL disabilities for the first year after a patient's cancer diagnosis. The first method is a standard rule learning algorithm; the second is that same algorithm additionally equipped with methods for reasoning with ontologies. The models showed that a patient's race, ethnicity, smoking preference, treatment plan and tumor characteristics including histology, staging, cancer site, and morphology were predictors for ADL performance levels one year after cancer diagnosis. The ontology-guided ML method was more accurate at predicting ADL performance levels (P < 0.1) than methods without ontologies. CONCLUSIONS: This study demonstrated that bio-ontologies can be harnessed to provide medical knowledge for ML algorithms. The presented method demonstrates that encoding specific types of hierarchical relationships to guide rule learning is possible, and can be extended to other types of semantic relationships present in biomedical ontologies. The ontology-guided ML method achieved better performance than the method without ontologies. The presented method can also be used to promote the effectiveness and efficiency of ML in healthcare, in which use of background knowledge and consistency with existing clinical expertise is critical.

A Comprehensive Multimorbidity Index for Predicting Mortality in Intensive Care Unit Patients.

BACKGROUND: Accurate prediction of mortality for patients admitted to the intensive care units (ICUs) is an important component of medical care. However, little is known about the role of multimorbidity in predicting end of life for high-risk and vulnerable patients. OBJECTIVE: The aim of the study was to derive and validate a multimorbidity risk model in an attempt to predict all-cause mortality at 6 and 12 months posthospital discharge. METHODS: This is a retrospective, observational, clinical cohort study. Data were collected on 442,692 ICU patients who received care through the Veterans Administration between January 2003 and December 2013. The primary outcome was all-cause mortality at 6 and 12 months posthospital discharge. We divided the data into derivation (80%) and validation (20%) sets. Using multivariable logistic regression models, we compared prognostic models based on age, principal diagnosis groups, physiological markers, immunosuppressants, comorbidity categories, and a newly developed multimorbidity index (MMI) based on 5695 comorbidities. The cross-validated area under the receiver operating characteristic curve (AUC) was used to report the accuracy of predicting all-cause mortality at 6 and 12 months of hospital discharge. RESULTS: The average age of patients was 68.87 years (standard deviation = 12.1), 95.9% were males, 44.9% were widowed, divorced, or separated. The relative order of accuracy in predicting mortality was the MMI (AUC = 0.84, CI = 0.83-0.84), VA Inpatient Evaluation Center index (AUC = 0.80, CI = 0.79-0.81), principal diagnosis groups (AUC = 0.74, CI = 0.73-0.74), comorbidities (AUC = 0.69, CI = 0.68-0.70), physiological markers (AUC = 0.65, CI = 0.64-0.65), age (AUC = 0.60, CI = 0.60-0.61),and immunosuppressant use (AUC = 0.59, CI = 0.58-0.59). CONCLUSIONS: The MMI improved the accuracy of predicting short- and long-term all-cause mortality for ICU patients. Further prospective studies are needed to validate the index in different clinical settings and test generalizability of results in patients outside the VA system of care.