Predictions of psychiatric hospitalization and emergency department utilization in commercially insured late middle-aged adults with depression

Predictions of psychiatric hospitalization and emergency department utilization in commercially insured late middle-aged adults with depression | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Predictions of psychiatric hospitalization and emergency department utilization in commercially insured late middle-aged adults with depression Wenna Xi, Lauren Evans, Yiyuan Wu, Min-hyung Kim, Arnab Ghosh, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4902124/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background : Late middle-aged adults with depression experience a high illness burden often resulting from multiple chronic conditions. Risk stratification using administrative claims data is a promising method to identify enrollees at highest risk of experiencing psychiatric hospitalization and emergency department (ED) utilization. The goal of this study was to develop and validate a non-proprietary model to identify late middle-aged adults with depression at risk of 1- and 2- year psychiatric hospitalization and ED utilization, with broad applicability to commercially insured adults. Methods : We conducted a retrospective cohort study on 71,682 commercially insured adults aged 55-64 years with depression diagnosis. National health insurance claims were used to capture demographics, health care utilization, and health status during the baseline year. Health status was captured using 70 chronic health conditions, and 46 mental health conditions. The four outcomes were one-year and two-year psychiatric hospitalization and ED utilization. Results : For 1-year risk of psychiatric hospitalization, the final model achieved an AUC of 0.806, with a sensitivity of ­­61% and a specificity of 85%. For 2-year risk of psychiatric hospitalization, the final model achieved an AUC of 0.781, with a sensitivity of 68% and a specificity of 75%. For 1-year risk of psychiatric ED utilization, the final model achieved an AUC of 0.727, with a sensitivity of 56% and specificity of 78%. For the 2-year risk of psychiatric ED utilization, the final model achieved an AUC of 0.748, with a sensitivity of 67% and a specificity of 72%. The final models for all four outcomes outperformed their corresponding models using only demographics and prior utilization. Conclusions: Predictive modeling using diagnoses and other demographic characteristics readily available in claims data can be used to identify late middle-aged adults with depression at risk of psychiatric hospitalization and ED utilization. Given the aging of the U.S. population and the significant burden of illness experienced by those with depression and multiple chronic conditions, this approach may assist policy makers and health care administrators in the identification of high-risk patients who may benefit from additional screening and care management interventions. insurance claims data psychiatric hospitalization psychiatric emergency department service utilization risk adjustment late middle-aged adults depression machine learning Figures Figure 1 Figure 2 Background Depression is associated with a number of adverse health outcomes and poorer patient experiences. Patients with depression have been shown to have higher health care costs after adjustment for chronic medical illness, 1 higher costs of medical services, 2 and a higher use of acute medical services such as emergency department use, 3 and hospitalization. 3 , 4 Much of the research involving psychiatric hospitalization or emergency department (ED) use in older adults involves cross-sectional descriptive analysis, such as the clinical descriptors and service use. 5 – 8 In these types of studies, we see that chronic conditions and medical comorbidity are common in older adults who have experienced psychiatric hospitalization or ED use. Prior research involving older adults reveals that depression or bipolar disorder is the most common reason for admission, often co-occurring with anxiety disorders or substance abuse, and that the presence dementia or cognitive impairment among elderly inpatients complicates diagnosis and treatment. 6 Prior research suggests that, nationally, about 0.6% of elderly persons with Medicare are hospitalized for mental illness, 5 with older patients having more medical comorbidity tending to be treated in general hospitals with specialty units. 5 In adult inpatient hospitalizations in those aged 45–64 years where the primary diagnosis is a mental health or substance use disorder, more than 90% have a co-occurring medical condition. 9 When it comes to ED utilization for psychiatric conditions or alcohol/substance use disorders, late middle-aged and older adults have relatively high rates, with important differences noted by sex. The rate of ED visits per 100,000 persons among adults aged 45 years and above for substance use disorders is 6,056 for males and 2,598 for females; for depression, anxiety and stress reactions, it is 7,021 for males and 11,965 for females; and for psychoses and bipolar disorder, it is 2,508 for males and 2,701 for females. 8 In addition to different patterns of psychiatric utilization based on sex and age, we also see different patterns of utilization based on primary payer. For example, schizophrenia is a much less common cause of psychiatric hospitalization in populations with private insurance, 7 , 10 and substance use is a much less common cause of ER visits for privately insured individuals. 8 This may be due to the difficulty in maintaining commercial insurance in those with significant medical comorbidities, and that Medicaid programs have historically provided a greater proportion of care to people with disabling psychiatric illness. 11 However, psychiatric hospitalization and ED utilization is still a concern among the privately insured, as between 14–18% of mental health or substance use disorder ED visits in 2006–2013 were among the commercially insured, 8 and in approximately 27% of inpatient stays in 2016 with a primary diagnosis of mental health or substance use disorder, the expected payer was private insurance. 12 There is strong interest in identifying patients likely to experience psychiatric hospitalization and ED utilization. Yet efforts to predict patients at risk for these outcomes have largely focused on relatively short prediction windows, such as the 30-day period following the index event 7 , 10 , 13 or within 90 days of admission, 13 rather than a first or future hospitalization/ED visit within the next year or longer. Mental health specialists and primary care providers have few tools to help identify at-risk patients. Predictive modeling of the first psychiatric hospitalization or ED visit informs treatment by clinicians, and presenting such models in a transparent, non-proprietary format extends their utility for other audiences such as researchers, administrators, and policy makers. Our goal in the present study was to develop and validate non-proprietary predictive models for 1-year and 2-year psychiatric hospitalization and ED utilization, separately, for late middle-aged adults with depression, using nationally representative commercial insurance claims data from Health Care Cost Institute (HCCI). Specifically, we developed and validated separate predictive models of risk of psychiatric hospitalization (1-year and 2-year risk) and psychiatric ED utilization (1-year and 2-year risk). We used the diagnostic categories used in the CMS HCC risk adjustment system, 14 together with the Psychiatric Case-Mix System (PsyCMS) developed in the Veterans Affairs health system, 15 as well as demographic characteristics, and prior healthcare utilization measures to capture the health status and prior healthcare utilization of the patient population in the baseline year. Given the lack of access to care in many rural areas and the complex ways that sex influences chronic and mental health conditions, we examined how the relationship between chronic and mental health conditions and psychiatric hospitalization/ED utilization varied by sex and rural/urban residence. Methods Data source This retrospective cohort study used commercial health insurance claims data from the Health Care Cost Institute (HCCI). 16 The HCCI data include nationwide de-identified claims of beneficiaries covered by four of the nation’s largest health insurers (i.e., Aetna, Humana, Kaiser Permanente, and UnitedHealthcare) in a manner compliant with the Health Insurance Portability and Accountability Act (HIPAA). This study was determined to be exempt by the Institutional Review Board of Weill Cornell Medicine. Establishment of the Study Cohort Enrollees who were aged 55–64 years with continuous insurance coverage for at least 36-months from January 2011 through December 2013 were considered for inclusion in the study sample. The continuous enrollment requirement was imposed to accurately capture enrollees’ medical history and risk of hospitalization in the study period. The 36 months of the study period were split into 1) Year 1 (the “baseline year” to capture demographics and medical history); and 2) Year 2 (to capture the 1-year risk of psychiatric hospitalization/ED visit); and 3) Years 2 and 3 combined (to capture the 2-year risk of psychiatric hospitalization/ED visit). Enrollees were excluded if they had a hospice or nursing home claim in the base year or if they did not have at least one medical claim during the baseline year. We used a validated method to identify patients with depression in administrative data. 17 , 18 We required at least one inpatient claim for depression, or two outpatient or two physician claims with a diagnosis of depression, or one outpatient or physician claim for depression plus at least one antidepressant medication fill during the baseline year. Depression diagnosis was defined using the following International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes: 296.20, 296.21, 296.22, 296.23, 296.24, 296.25, 296.26, 296.30, 296.31, 296.32, 296.33, 296.34, 296.35, 296.36, 300.4, and 311. Antidepressant medication fills were identified using the HEDIS National Drug Code (NDC) list for Antidepressant Medication Management, produced by the National Committee for Quality Assurance that was in effect during the enrollee’s baseline year. Measures All candidate predictors were binary and were measured in the baseline year only. Candidate predictors included demographic characteristics, prior healthcare utilization measures, and indicators of presence of selected medical and mental health diagnoses. Details of each of these measures are described below. Demographic characteristics Demographic variables included in the models were sex and urbanicity. Sex (male vs. female) was included because of the gender differences in the course and development of medical and mental health conditions. Urbanicity was included as a binary variable to indicate whether the individual resided in a core-based statistical area (CBSA), as defined by the U.S. White House Office of Management and Budget and using population counts collected in the decennial Census. 19 In this approach, enrollees are considered as living in an urban setting if they reside in an area with a population of at least 50,000. Healthcare Utilization Measures Inclusion of prior healthcare utilization measures can improve model performance. 21 We included four dichotomous variables to capture prior healthcare utilization in the baseline year: all-cause hospitalization, hospitalization for a psychiatric condition, all-cause ED visit, and ED visit for a psychiatric condition. Hierarchical Condition Categories (HCCs) We used the 70 condition categories defined in Version 12 of the CMS HCC model, which was in use for claims incurred during the study observation window (See Supplementary Material Table S1 ). 22 HCC risk adjustment systems have been used to adjust capitation payments to Medicare Advantage Plans for the health expenditure risk for their enrollees. 14 It has also been used in health services research to predict outcomes such as hospitalization risk 24 , 25 and other utilization outcomes such as emergency department use, hospital readmission, and the likelihood of being a high cost enrollee. 25 For related condition categories in HCCs, hierarchies are imposed so that a patient with multiple related conditions is coded for only the most severe one. The unrelated condition categories in the HCCs are additive; an individual may be coded for none, one, or multiple HCCs. 23 PsyCMS psychiatric condition categories Many mental health and substance use conditions do not map to conditions specified in the HCC model. 26 In order to more adequately capture the mental health and substance use conditions not captured in the CMS HCC V12 model, we incorporated the 46 mental health and substance use condition categories defined in the PsyCMS Case-Mix System (See Supplementary Material Table S2). This model was developed using a national sample of Department of Veterans Affairs (VA) patients with mental health and substance use disorders. 15 The PsyCMS system also imposes hierarchies, based on the clinical assessment of severity, medical diagnostic criteria, and greater specificity, to reduce overlap among closely related diagnosis codes, 15 which allows us to account for both multimorbidity and the overlap among conditions. Mental Health Outcomes We modeled 4 binary outcomes reflecting adverse mental health outcomes in the 12- and 24-month prediction period: 1- and 2-year risk of psychiatric hospitalization, and 1- and 2-year risk of ED utilization for a psychiatric condition. Psychiatric hospitalization or ED visit were identified if the primary diagnosis for the hospitalization or ED visit was one of the psychiatric condition codes identified by the psychiatric codes used in previous studies. 28 Statistical Analysis Our predictive modeling process involved fitting a series of models, and then, for each outcome, identifying the one with the best performance as the final model. The first set of models (Models 1 through 4) were logistic regression models, with different combinations of predictors to evaluate the relative importance of each group of variables. Predictors of Model 1 were demographics (sex and urbanicity) and healthcare utilization measures. Predictors of Model 2 were demographics and HCC conditions. Predictors of Model 3 were demographics and PsyCMS conditions. Predictors of Model 4 were demographics, healthcare utilization measures, HCC conditions, and PsyCMS conditions. The second set of models (Models 5 through 7) were machine learning models using all predictors (demographics, healthcare utilization measures, HCC conditions, and PsyCMS conditions). In addition, Model 5, which utilized logistic regression with Least Absolute Shrinkage and Selection Operator (LASSO) penalty, 29 also included two-way interactions of all predictors with sex and with urbancity. Model 6 was a random forests (RF) model 30 ; and Model 7 was a gradient boosting machine (GBM) model. 31 We chose two classes of ML algorithms: a regression-based method (LASSO) that models additive effects of predictors along with simple two-way interactions of predictors with sex and a marker for rural/urban residence; and methods based on decision trees predictors (RF and GBM) which model complex higher order interactions involving predictors. The LASSO model is a regression-based method that, by applying shrinkage to the coefficients in the model, achieves better prediction accuracy due to reduced variance. 32 RF is an ensemble of decision trees that fits uncorrelated decision trees using bootstrapped samples and then combines them using a technique called bagging. 34 GBM is also an ensemble learning method which uses boosting to produce a complex aggregated model by combining simple prediction models. 33 We used a 2:1 ratio for the train-test split: 2/3 of the sample was randomly selected as the training set for model development, whereas the remaining 1/3 was used as the test set to evaluate the model performance. To train the models on balanced data, the training set was downsampled by randomly selecting nonevents to match the events in a 1:1 ratio for each outcome. Before fitting each model, predictors with a frequency ratio (i.e., the ratio of the frequency of the most common category and the other category) > 20 were dropped because of the low information content. All models were trained on the training set, with a five-fold cross-validation to tune the parameters for machine learning models, and model performance metrics (sensitivity, specificity, and area under the receiver operating characteristics curve (AUC)) were estimated on the test set. The sensitivity and specificity of our models were calculated by choosing a threshold for the risk prediction functions that maximized the Youden index. 35 All logistic regression models were examined for calibration (agreement between predicted and observed frequencies) using the Brier score and visual inspection, and then, if needed, re-calibrated coefficients were obtained on the training data. Results Overall, 71,682 adults aged between 55 to 64 years with a depression diagnosis and with at least 36 months of continuous enrollment met our inclusion criteria. After downsampling, for the 1-year risk of psychiatric hospitalization analysis, the training set contained 1,128 patients and the test set contained 23,893 patients. For the 2-year risk of psychiatric hospitalization analysis, the training set contained 1,904 patients and the test set contained 23,893 patients. For the 1-year risk of psychiatric ED visit analysis, the training set contained 1,306 patients and the test set contained 23,893 patients. For the 2-year risk of psychiatric ED visit analysis, the training set contained 2,282 patients and the test set contained 23,893 patients. Among the 71,682 patients included in the study cohort, 1.18% (n = 845) experienced a psychiatric hospitalization in the 1-year prediction window, 1.99% (n = 1,427) experienced a psychiatric hospitalization in the 2-year prediction window, 1.37% (n = 979) utilized ED services for a psychiatric condition within the 1-year prediction window, and 2.39% (n = 1,711) utilized ED services for a psychiatric condition in the 2-year prediction window. Approximately 70% of the sample was female, and all were aged 55–64 years during the observation period. Nearly 90% resided in a metropolitan core-based statistical area. The model performance (AUC) of the seven models for each of the four outcomes are summarized in Table 1 . The ROC curves of the logistic regression models for the four outcomes are presented in Figs. 1 a-d, respectively. Table 1 Comparison of Prospective Model Performance Using Area under ROC Curve (AUC) for 1- and 2-year risk of Psychiatric Hospitalization and Emergency Department Use for Psychiatric Conditions Model Variables used in Alternative Hospitalization Prediction Models* Outcome† Model 1 = Dem. and prior utilization Model 2 = Dem. and HCC Model 3 = Dem. and PsyCMS Model 4 = Dem., HCC and PsyCMS Model 5 = Final model ‡ utilizing machine learning – logistic regression with Least Absolute Shrinkage and Selection Operator (LASSO) penalty Model 6 utilizing machine learning – random forests Model 7 utilizing machine learning – gradient boosting 1-year risk of psychiatric hospitalization .715 .711 .781 .792 .806 .727 .756 2-year risk of psychiatric hospitalization .694 .694 .757 .777 .781 .715 .749 1-year risk of psychiatric emergency department use .648 .662 .713 .725 .727 .677 .699 2-year risk of psychiatric emergency department use .673 .679 .717 .737 .748 .692 .727 *All predictors and risk data sources were derived from year 1 † 1-year prospective outcome data was measured using claims data from year 2, and 2-year prospective outcome data was measured using claims data from years 2 and 3 combined ‡ The Final model included all variables specified in other models plus interaction terms, after applying lasso and filtering Dem. indicates demographics; HCC indicates Hierarchical Condition Categories; PsyCMS indicates Psychiatric Case-Mix System conditions For each of the four outcomes, our best performing model was Model 5 (hereinafter referred to as the Final Model), the machine learning approach of logistic regression with LASSO penalty, using demographic characteristics, healthcare utilization measures, HCC conditions, PsyCMS conditions, and interaction effects between sex and all other variables and between urbanicty and all other variables as potential predictors. For the 1-year risk of psychiatric hospitalization, our final model achieved an AUC of 0.806, with a sensitivity of 61% and a specificity of 85% obtained under the optimal threshold 0.56. For the 2-year risk of psychiatric hospitalization, our final model achieved an AUC of 0.781, with a sensitivity 68% and a specificity of 75% under the optimal threshold 0.43. For the 1-year risk of psychiatric ED visit, our final model achieved an AUC of 0.727, with a sensitivity of 56% and a specificity of 78% under the optimal threshold 0.53. For the 2-year risk of psychiatric ED visit, our final model achieved an AUC of 0.748, with a sensitivity of 67% and a specificity of 72% using the optimal threshold of 0.47. Our final models (Model 5), logistic regressions with LASSO penalty, allowed us to identify main effects and interaction effects that affected the odds of experiencing psychiatric hospitalization or ED utilization. Figures 2 a-d present risk factors selected for each of the four outcomes, along with the odds ratios and 95% confidence intervals associated with these risk factors in our final models. The following factors increased the odds of experiencing all four outcomes: prior all-cause ED utilization, the PsyCMS category of anxiety not otherwise specified, the PsyCMS category of bipolar disorders, and the HCC category of major depressive and bipolar and paranoid disorders. The following factors increased the odds of three of the outcomes (i.e., 1-year psychiatric hospitalization, 2-year psychiatric hospitalization, and 2-year psychiatric ED utilization): prior psychiatric hospitalization, and the HCC category of drug/alcohol dependence. The PsyCMS category of severe Major Depressive Disorder emerged as an important risk factor in the final models for three of the four outcomes (specifically, in the final models for 1-year psychiatric hospitalization, 2-year psychiatric hospitalization, and 1-year psychiatric ED utilization). Organic Other conditions (a group of conditions specified in the PsyCMS system that include conditions such as alcohol-induced mental disorders, delusions, withdrawal and withdrawal delirium, various drug-induced mental disorders, various other organic psychotic conditions, and other nonpsychotic mental disorders due to organic brain damage) were an important risk factor in the final models for both the 1-year and 2-year psychiatric hospitalization. A diagnosis of unspecified psychosis (the condition category of Psychosis NOS in the PsyCMS system) was an important factor in the final models for both 1-year and 2-year psychiatric ED utilization. Certain mental health conditions and interactions emerged as important factors in the final models for only one of the four outcomes: the PsyCMS category of major psychiatric disorder in remission, and the interaction between Generalized Anxiety Disorder and sex were important in the 1-year psychiatric hospitalization final model; the PsyCMS category of Depression not otherwise specified, and the interaction between the PsyCMS category of Major Depressive Disorder (Single Episode) and sex was important in the 2-year psychiatric hospitalization final model; and the PsyCMS category of panic/agoraphobia emerged as an important risk factor in the 2-year psychiatric ED utilization final model. Figure 2 a: Selected risk factors for 1-year risk of psychiatric hospitalization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI. Risk factors above the blue dashed line correspond to the interaction between Generalized Anxiety and Sex. Marginal estimates of risk are presented for the four categories of the interaction effect adjusting for other risk factors in the model. Figure 2 b: Selected risk factors for 2-year risk of psychiatric hospitalization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI. Risk factors above the blue dashed line correspond to the interaction between Major Depressive Disorder (Single Episode) and Sex. Marginal estimates of risk are presented for the four categories of the interaction effect adjusting for other risk factors in the model. Importantly, we were able to identify medical conditions that increased the odds of the psychiatric outcomes modeled in this paper. The HCC category of chronic obstructive pulmonary disease (COPD) emerged as a risk factor in both the 1-and 2-year final models for psychiatric hospitalization; and the HCC categories of polyneuropathy and congestive heart failure were identified as important factors in the 2-year psychiatric ED utilization final model. Discussion In this study, we developed custom calibrated models to predict psychiatric hospitalization and ED utilization for psychiatric conditions among commercially insured late middle-aged adults with depression. Our models involved the use of a large set of clinical, demographic, and prior healthcare utilization variables to provide a rich description of the patients that can be used for identifying patients for care management. One benefit of our approach is that it can be easily modified to include other health conditions or demographic characteristics, such as race and household income, when available. We found that it was preferable to use a more complete method for the classification of medical and mental health and behavioral health factors, which is consistent with prior studies. 23 , 36 , 37 We found that the inclusion of the psychiatric condition categories, in addition to the HCC conditions, enhanced the performance of our models, thus allowing us to better identify high-risk patients. Our best performing models for each of the four outcomes modeled (1- and 2-year risk of psychiatric hospitalization, and 1-and 2-year risk of ED utilization for psychiatric conditions) used the machine learning approach of logistic regression with LASSO penalty. This approach outperformed the other two candidate machine learning approaches of RF and GBM, indicating that a regression-based model with additive effects of predictors along with hypothesized two-way interaction effects between all predictors with sex and urbanicity can sufficiently capture the variability in the outcomes, without the need to include complex higher-order interaction terms by using RF or GBM. In addition, logistic regression with LASSO penalty produces findings that are transparent with interpretable model parameters and the approach allows for the exploration of simple interactions among model variables. Unsurprisingly, several psychiatric comorbidities emerged as significant risk factors across several of the outcomes measured (e.g., severe major depressive, bipolar and paranoid disorders, anxiety, and drug/alcohol dependence, psychoses, psychoses and organic other conditions). We also found that interaction effects were observed between sex and having a single episode of Major Depressive Disorder, and between sex and Generalized Anxiety Disorder, indicating that females with these conditions are at higher risk. Further, we identified several medical comorbidities that increased the risk of psychiatric hospitalization and ED utilization, namely COPD (for the two psychiatric hospitalization outcomes), vascular disease (for 2-year psychiatric hospitalization), and polyneuropathy and congestive heart failure (for 2-year psychiatric ED utilization). Strengths of our study include the use of insurance claims from multiple large U.S. insurers rather than from a single insurer or geographic region, allowing for wider applicability to commercially insured late-middle aged adults. In addition, the calculation of our risk prediction functions is easily replicable, as they rely solely on claims and demographic characteristics and do not rely on additional collection of clinical assessments or survey-based measures. Assessments and surveys may offer important prognostic variables that are not available through claims data, however availability and accessibility may depend on the standardization and reach of such assessments and the information technology environment 38 and survey-based measures often suffer from significant nonresponse bias where nonrespondents are often at higher risk of health status decline and subsequent utilization. 24 There are several limitations of the current study. First, we did not have access to certain demographic characteristics (e.g., race, ethnicity) to protect patient privacy. Second, the chronic illnesses in the baseline year were identified using diagnosis codes in claims for billing purposes; we acknowledge that, in claims data, enrollees may be miscoded because of coding errors and billing purposes. Third, our inclusion criterion of at least 36 months of continuous enrollment allowed us to predict enrollees’ health outcomes over a longer time window, but might limit the generalizability of our results to the general patient population, since patients with continuous enrollment might have more stable jobs and share a different demographic and health profile. Fourth, our study focused on adults with commercial insurance, the largest payer type in the US, nevertheless, our findings may not be applicable to other payer types, such as Medicare, Medicaid, and uninsured. Conclusions We demonstrated that our predictive modeling approach using demographic and clinical characteristics readily available in claims data can be used to identify older adults with depression who may be at an increased risk of experiencing psychiatric hospitalization/ED utilization. Given the aging of the U.S. population and the significant burden of illness experienced by those with depression and multiple chronic conditions, this approach may assist policy makers and health care administrators in the identification of high-risk patients who may benefit from improved care delivery and monitoring. Declarations Ethics approval and consent to participate The Institutional Review Board of Weill Cornell Medicine approved this study. The informed consent was waived by Weill Cornell Medicine Institutional Review Board because it involved secondary data analysis using deidentified data. All methods were carried out in accordance with relevant guidelines and regulations. Consent for publication Not applicable Availability of data and materials The data that support the findings of this study are available from Health Care Cost Institute (HCCI) but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of HCCI. Competing interests The authors declare that they have no competing interests Funding This work was supported by the National Institutes of Health (grant numbers P50MH113838, R01MH105384, T32MH073553, and K99MH130713). Authors' contributions Study Concept and Design: SB, JP, GA, and LE Acquisition of Data: SB and JP Analysis of Data: YW and MK Interpretation of Data: LE, WX, and AG Drafting and Revision of the Manuscript: LE, WX, and AG Acknowledgements Not applicable References Katon WJ, Lin E, Russo J, Unutzer J. Increased medical costs of a population-based sample of depressed elderly patients. Arch Gen Psychiatry. 2003;60(9):897-903. Unutzer J, Patrick DL, Simon G, et al. Depressive symptoms and the cost of health services in HMO patients aged 65 years and older. A 4-year prospective study. Jama. 1997;277(20):1618-1623. Himelhoch S, Weller WE, Wu AW, Anderson GF, Cooper LA. Chronic medical illness, depression, and use of acute medical services among Medicare beneficiaries. Med Care. 2004;42(6):512-521. Huang BY, Cornoni-Huntley J, Hays JC, Huntley RR, Galanos AN, Blazer DG. Impact of depressive symptoms on hospitalization risk in community-dwelling older persons. J Am Geriatr Soc. 2000;48(10):1279-1284. Ettner SL. The setting of psychiatric care for medicare recipients in general hospitals with specialty units. Psychiatric services (Washington, DC). 2001;52(2):237-239. Prince JD, Akincigil A, Kalay E, et al. Psychiatric rehospitalization among elderly persons in the United States. Psychiatric services (Washington, DC). 2008;59(9):1038-1045. Heslin KC (AHRQ), Weiss AJ (Truven Health Analytics). Hospital Readmissions Involving Psychiatric Disorders, 2012. HCUP Statistical Brief #189. May 2015. Agency for Healthcare Research and Quality, Rockville, MD. http://www.hcup-us.ahrq.gov/reports/statbriefs/sb189-Hospital-Readmissions-Psychiatric-Disorders-2012.pdf . Weiss AJ, Barrett ML, Heslin KC, Stocks C. Trends in Emergency Department Visits Involving Mental and Substance Use Disorders, 2006–2013.HCUP Statistical Brief #216. December 2016. Agency for Healthcare Research and Quality, Rockville, MD. Accessed December 23, 2019. Available at: http://www.hcup-us.ahrq.gov/reports/statbriefs/sb216-Mental-Substance-Use-Disorder-ED-Visit-Trends.pdf . . Owens PL, Heslin KC, Fingar KR, Weiss AJ. Co-occurrence of Physical Health Conditions and Mental Health and Substance Use Conditions Among Adult Inpatient Stays, 2010 Versus 2014. HCUP Statistical Brief #240. June 2018. Agency for Healthcare Research and Quality, Rockville, MD. www.hcup-us.ahrq.gov/reports/statbriefs/sb240-Co-occurring-Physical-Mental-Substance-Conditions-Hospital-Stays.pdf . Strom JB, Kramer DB, Wang Y, et al. Short-term rehospitalization across the spectrum of age and insurance types in the United States. PLoS One. 2017;12(7):e0180767. Tsai J, Rosenheck RA, Culhane DP, Artiga S. Medicaid expansion: chronically homeless adults will need targeted enrollment and access to a broad range of services. Health Aff (Millwood). 2013;32(9):1552-1559. Owens PL, Fingar KR, McDermott KW, Muhuri PK, Heslin KC. Inpatient Stays Involving Mental and Substance Use Disorders, 2016: Statistical Brief #249. In: Healthcare Cost and Utilization Project (HCUP) Statistical Briefs. Rockville (MD): Agency for Healthcare Research and Quality (US); 2006. Perlman CM, Hirdes JP, Vigod S. Psychiatric Rehospitalization: Development of a Person-Level Indicator for Care Planning and Quality Assurance. Prim Care Companion CNS Disord. 2015;17(4). Pope GC, Kautter J, Ellis RP, et al. Risk adjustment of Medicare capitation payments using the CMS-HCC model. Health care financing review. 2004;25(4):119-141. Sloan KL, Montez-Rath ME, Spiro A, 3rd, et al. Development and validation of a psychiatric case-mix system. Med Care. 2006;44(6):568-580. Newman D, Herrera CN, Parente ST. Overcoming barriers to a research-ready national commercial claims database. Am J Manag Care. 2014;20(11 Spec No. 17):eSP25-30. Townsend L, Walkup JT, Crystal S, Olfson M. A systematic review of validated methods for identifying depression using administrative data. Pharmacoepidemiol Drug Saf. 2012;21 Suppl 1:163-173. Fiest KM, Jette N, Quan H, et al. Systematic review and assessment of validated case definitions for depression in administrative data. BMC Psychiatry. 2014;14:289. Hall SA, Kaufman JS, Ricketts TC. Defining urban and rural areas in U.S. epidemiologic studies. J Urban Health. 2006;83(2):162-175. Everson J, Hollingsworth JM, Adler-Milstein J. Comparing methods of grouping hospitals. Health Serv Res. 2019;54(5):1090-1098. Lemke KW, Weiner JP, Clark JM. Development and validation of a model for predicting inpatient hospitalization. Med Care. 2012;50(2):131-139. Evaluation of the CMS-HCC Risk Adjustment Model: Final Report. Authors: Pope GC, Kautter J, Ingber MJ, Freeman S, Sekar R, Newhart C. Federal Project Officer: Melissa A. Evans, PhD. RTI International. CMS Contract No. HHSM-500-2005-00029I TO 0006. March 2011. Rosen AK, Loveland SA, Anderson JJ, Hankin CS, Breckenridge JN, Berlowitz DR. Diagnostic cost groups (DCGs) and concurrent utilization among patients with substance abuse disorders. Health services research. 2002;37(4):1079-1103. Mosley DG, Peterson E, Martin DC. Do hierarchical condition category model scores predict hospitalization risk in newly enrolled Medicare advantage participants as well as probability of repeated admission scores? J Am Geriatr Soc. 2009;57(12):2306-2310. Haas LR, Takahashi PY, Shah ND, et al. Risk-stratification methods for identifying patients for care coordination. Am J Manag Care. 2013;19(9):725-732. Wagner TH, Almenoff P, Francis J, Jacobs J, Pal Chee C. Assessment of the Medicare Advantage Risk Adjustment Model for Measuring Veterans Affairs Hospital Performance. JAMA Netw Open. 2018;1(8):e185993. Kautter J, Pope GC, Ingber M, et al. The HHS-HCC risk adjustment model for individual and small group markets under the Affordable Care Act. Medicare Medicaid Res Rev. 2014;4(3). Johnston KJ, Allen L, Melanson TA, Pitts SR. A "Patch" to the NYU Emergency Department Visit Algorithm. Health Serv Res. 2017;52(4):1264-1276. Hastie T, Tibshirani R, Friedman J. Linear Methods for Regression. In: The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed. New York: Springer; 2009. Hastie T, Tibshirani R, Friedman J. Random Forests. In: The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed. New York: Springer; 2009. Hastie T, Tibshirani R, Friedman J. Boosting and Additive Trees. In: The Elements of Statistical Learning: Data Mining, Inference, and Prediction. 2nd ed. New York: Springer; 2009. R T. Regression shrinkage and selection via the Lasso Journal of the Royal Statistical Society. 1996;58(1):267-288. Friedman JH. Greedy function approximation: A gradient boosting machine. Ann Statist. 2001;29(5):1189-1232. Breiman L. Random Forests. Machine Learning. 2001;45(1):5-32. Fluss R, Faraggi D, Reiser B. Estimation of the Youden Index and its associated cutoff point. Biom J. 2005;47(4):458-472. Shrestha A, Bergquist S, Montz E, Rose S. Mental Health Risk Adjustment with Clinical Categories and Machine Learning. Health Serv Res. 2018;53 Suppl 1:3189-3206. Ettner SL, Frank RG, McGuire TG, Hermann RC. Risk adjustment alternatives in paying for behavioral health care under Medicaid. Health services research. 2001;36(4):793-811. Kinosian B, Wieland D, Gu X, Stallard E, Phibbs CS, Intrator O. Validation of the JEN frailty index in the National Long-Term Care Survey community population: identifying functionally impaired older adults from claims data. BMC Health Serv Res. 2018;18(1):908. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4902124","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":339690982,"identity":"b82aa787-6367-4522-a7db-c355236af3db","order_by":0,"name":"Wenna Xi","email":"","orcid":"","institution":"Weill Cornell Medicine","correspondingAuthor":false,"prefix":"","firstName":"Wenna","middleName":"","lastName":"Xi","suffix":""},{"id":339690983,"identity":"dea98247-64c8-47a6-9ad2-a26227011917","order_by":1,"name":"Lauren Evans","email":"","orcid":"","institution":"VNS Health","correspondingAuthor":false,"prefix":"","firstName":"Lauren","middleName":"","lastName":"Evans","suffix":""},{"id":339690984,"identity":"2b5908e3-2c1a-4bae-a91a-04276c68b04b","order_by":2,"name":"Yiyuan Wu","email":"","orcid":"","institution":"Weill Cornell Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yiyuan","middleName":"","lastName":"Wu","suffix":""},{"id":339690985,"identity":"4b6f9d57-a71d-440b-9406-4fd8e7fe4588","order_by":3,"name":"Min-hyung Kim","email":"","orcid":"","institution":"Seoul National University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Min-hyung","middleName":"","lastName":"Kim","suffix":""},{"id":339690986,"identity":"c143b8ee-afb4-4026-a387-1dda53154685","order_by":4,"name":"Arnab Ghosh","email":"","orcid":"","institution":"Weill Cornell Medicine","correspondingAuthor":false,"prefix":"","firstName":"Arnab","middleName":"","lastName":"Ghosh","suffix":""},{"id":339690987,"identity":"42136ac8-af9b-4db5-9247-125f3f6623ee","order_by":5,"name":"George Alexopoulos","email":"","orcid":"","institution":"Weill Cornell Medicine","correspondingAuthor":false,"prefix":"","firstName":"George","middleName":"","lastName":"Alexopoulos","suffix":""},{"id":339690988,"identity":"1c7c9d8c-14c4-4889-9e86-5b9797130bed","order_by":6,"name":"Jyotishman Pathak","email":"","orcid":"","institution":"Weill Cornell Medicine","correspondingAuthor":false,"prefix":"","firstName":"Jyotishman","middleName":"","lastName":"Pathak","suffix":""},{"id":339690991,"identity":"66bc7b61-7cb6-4458-b4e3-024364f23774","order_by":7,"name":"Samprit Banerjee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/ElEQVRIiWNgGAWjYFACHsYHCRU2ckCWAUSAnbmBgYENrxZmgw9n0owRWpgZCWphk5zZdjixgWgt8u1nD0jzsKWlr21v3sDwo+JeHj9Iy4eywzi1GJzJSzDm4bHJ3XbmWAFjz5niYslmxgbGGefwaGHIMUjmkUjL3XYjx4CZsS0hccNhxgZm3jbcWuT73xgc5jE4nG52/w1Ey36Qlr94tDDcyDFsnJFwOMHsBg/UFqBfgAw8DrvxLpnhw4E0w21n0goO9pxJSJwBtOVgz7l0PA7LPf4j8Z+NvNnxwxsf/KhISOxvbz744EeZNW6HIYMDGIxRMApGwSgYBeQBANEhXFEXYebzAAAAAElFTkSuQmCC","orcid":"","institution":"Weill Cornell Medicine","correspondingAuthor":true,"prefix":"","firstName":"Samprit","middleName":"","lastName":"Banerjee","suffix":""}],"badges":[],"createdAt":"2024-08-12 16:59:58","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4902124/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4902124/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":66124204,"identity":"3a3135ea-8903-4b84-908a-519bbf1a13cc","added_by":"auto","created_at":"2024-10-08 02:34:24","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2624609,"visible":true,"origin":"","legend":"\u003cp\u003e(a): Receiver operating characteristic (ROC) curves comparing the accuracy of the five logistic regression models in predicting 1-year risk of psychiatric hospitalization\u003c/p\u003e\n\u003cp\u003e(b): Receiver operating characteristic (ROC) curves comparing the accuracy of the five logistic regression models in predicting 2-year risk of psychiatric hospitalization\u003c/p\u003e\n\u003cp\u003e(c): Receiver operating characteristic (ROC) curves comparing the accuracy of the five logistic regression models in predicting 1-year risk of emergency department use for psychiatric conditions\u003c/p\u003e\n\u003cp\u003e(d): Receiver operating characteristic (ROC) curves comparing the accuracy of the five logistic regression models in predicting 2-year risk of emergency department use for psychiatric conditions\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4902124/v1/ca1cee64c9183efb95594608.png"},{"id":66124205,"identity":"b66350c6-a4c9-4054-8e49-ccdaf3d5c63f","added_by":"auto","created_at":"2024-10-08 02:34:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1691537,"visible":true,"origin":"","legend":"\u003cp\u003ea: Selected risk factors for 1-year risk of psychiatric hospitalization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI. Risk factors above the blue dashed line correspond to the interaction between Generalized Anxiety and Sex. Marginal estimates of risk are presented for the four categories of the interaction effect adjusting for other risk factors in the model.\u003c/p\u003e\n\u003cp\u003eb: Selected risk factors for 2-year risk of psychiatric hospitalization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI. Risk factors above the blue dashed line correspond to the interaction between Major Depressive Disorder (Single Episode) and Sex. Marginal estimates of risk are presented for the four categories of the interaction effect adjusting for other risk factors in the model.\u003c/p\u003e\n\u003cp\u003ec: Selected risk factors for 1-year risk of psychiatric ED utilization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI.\u003c/p\u003e\n\u003cp\u003ed: Selected risk factors for 2-year risk of psychiatric ED utilization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4902124/v1/950059b73e3a503cd6476240.png"},{"id":92837649,"identity":"7593dc84-d924-4fff-8d81-956454679ab1","added_by":"auto","created_at":"2025-10-06 08:09:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4684836,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4902124/v1/91964fdd-a7f8-43a1-a3f6-4e4f9e4310d2.pdf"},{"id":66124207,"identity":"99dfcb3a-9d4c-4a40-aad3-d9bc92fcf7da","added_by":"auto","created_at":"2024-10-08 02:34:27","extension":"docx","order_by":11,"title":"","display":"","copyAsset":false,"role":"supplement","size":21845,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-4902124/v1/ba1a457a9e46fe75b4f54cb1.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Predictions of psychiatric hospitalization and emergency department utilization in commercially insured late middle-aged adults with depression","fulltext":[{"header":"Background","content":"\u003cp\u003eDepression is associated with a number of adverse health outcomes and poorer patient experiences. Patients with depression have been shown to have higher health care costs after adjustment for chronic medical illness,\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e higher costs of medical services,\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e and a higher use of acute medical services such as emergency department use,\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e and hospitalization.\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e Much of the research involving psychiatric hospitalization or emergency department (ED) use in older adults involves cross-sectional descriptive analysis, such as the clinical descriptors and service use.\u003csup\u003e\u003cspan additionalcitationids=\"CR6 CR7\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e In these types of studies, we see that chronic conditions and medical comorbidity are common in older adults who have experienced psychiatric hospitalization or ED use.\u003c/p\u003e \u003cp\u003ePrior research involving older adults reveals that depression or bipolar disorder is the most common reason for admission, often co-occurring with anxiety disorders or substance abuse, and that the presence dementia or cognitive impairment among elderly inpatients complicates diagnosis and treatment.\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e Prior research suggests that, nationally, about 0.6% of elderly persons with Medicare are hospitalized for mental illness,\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e with older patients having more medical comorbidity tending to be treated in general hospitals with specialty units.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e In adult inpatient hospitalizations in those aged 45\u0026ndash;64 years where the primary diagnosis is a mental health or substance use disorder, more than 90% have a co-occurring medical condition.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eWhen it comes to ED utilization for psychiatric conditions or alcohol/substance use disorders, late middle-aged and older adults have relatively high rates, with important differences noted by sex. The rate of ED visits per 100,000 persons among adults aged 45 years and above for substance use disorders is 6,056 for males and 2,598 for females; for depression, anxiety and stress reactions, it is 7,021 for males and 11,965 for females; and for psychoses and bipolar disorder, it is 2,508 for males and 2,701 for females.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eIn addition to different patterns of psychiatric utilization based on sex and age, we also see different patterns of utilization based on primary payer. For example, schizophrenia is a much less common cause of psychiatric hospitalization in populations with private insurance,\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e and substance use is a much less common cause of ER visits for privately insured individuals.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e This may be due to the difficulty in maintaining commercial insurance in those with significant medical comorbidities, and that Medicaid programs have historically provided a greater proportion of care to people with disabling psychiatric illness.\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e However, psychiatric hospitalization and ED utilization is still a concern among the privately insured, as between 14\u0026ndash;18% of mental health or substance use disorder ED visits in 2006\u0026ndash;2013 were among the commercially insured,\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e and in approximately 27% of inpatient stays in 2016 with a primary diagnosis of mental health or substance use disorder, the expected payer was private insurance.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThere is strong interest in identifying patients likely to experience psychiatric hospitalization and ED utilization. Yet efforts to predict patients at risk for these outcomes have largely focused on relatively short prediction windows, such as the 30-day period following the index event\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e or within 90 days of admission,\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e rather than a first or future hospitalization/ED visit within the next year or longer. Mental health specialists and primary care providers have few tools to help identify at-risk patients. Predictive modeling of the first psychiatric hospitalization or ED visit informs treatment by clinicians, and presenting such models in a transparent, non-proprietary format extends their utility for other audiences such as researchers, administrators, and policy makers.\u003c/p\u003e \u003cp\u003e Our goal in the present study was to develop and validate non-proprietary predictive models for 1-year and 2-year psychiatric hospitalization and ED utilization, separately, for late middle-aged adults with depression, using nationally representative commercial insurance claims data from Health Care Cost Institute (HCCI). Specifically, we developed and validated separate predictive models of risk of psychiatric hospitalization (1-year and 2-year risk) and psychiatric ED utilization (1-year and 2-year risk). We used the diagnostic categories used in the CMS HCC risk adjustment system,\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e together with the Psychiatric Case-Mix System (PsyCMS) developed in the Veterans Affairs health system,\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e as well as demographic characteristics, and prior healthcare utilization measures to capture the health status and prior healthcare utilization of the patient population in the baseline year. Given the lack of access to care in many rural areas and the complex ways that sex influences chronic and mental health conditions, we examined how the relationship between chronic and mental health conditions and psychiatric hospitalization/ED utilization varied by sex and rural/urban residence.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData source\u003c/h2\u003e \u003cp\u003eThis retrospective cohort study used commercial health insurance claims data from the Health Care Cost Institute (HCCI).\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e The HCCI data include nationwide de-identified claims of beneficiaries covered by four of the nation\u0026rsquo;s largest health insurers (i.e., Aetna, Humana, Kaiser Permanente, and UnitedHealthcare) in a manner compliant with the Health Insurance Portability and Accountability Act (HIPAA). This study was determined to be exempt by the Institutional Review Board of Weill Cornell Medicine.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eEstablishment of the Study Cohort\u003c/h2\u003e \u003cp\u003eEnrollees who were aged 55\u0026ndash;64 years with continuous insurance coverage for at least 36-months from January 2011 through December 2013 were considered for inclusion in the study sample. The continuous enrollment requirement was imposed to accurately capture enrollees\u0026rsquo; medical history and risk of hospitalization in the study period. The 36 months of the study period were split into 1) Year 1 (the \u0026ldquo;baseline year\u0026rdquo; to capture demographics and medical history); and 2) Year 2 (to capture the 1-year risk of psychiatric hospitalization/ED visit); and 3) Years 2 and 3 combined (to capture the 2-year risk of psychiatric hospitalization/ED visit). Enrollees were excluded if they had a hospice or nursing home claim in the base year or if they did not have at least one medical claim during the baseline year.\u003c/p\u003e \u003cp\u003eWe used a validated method to identify patients with depression in administrative data.\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e We required at least one inpatient claim for depression, or two outpatient or two physician claims with a diagnosis of depression, or one outpatient or physician claim for depression plus at least one antidepressant medication fill during the baseline year. Depression diagnosis was defined using the following International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes: 296.20, 296.21, 296.22, 296.23, 296.24, 296.25, 296.26, 296.30, 296.31, 296.32, 296.33, 296.34, 296.35, 296.36, 300.4, and 311. Antidepressant medication fills were identified using the HEDIS National Drug Code (NDC) list for Antidepressant Medication Management, produced by the National Committee for Quality Assurance that was in effect during the enrollee\u0026rsquo;s baseline year.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eMeasures\u003c/h2\u003e \u003cp\u003eAll candidate predictors were binary and were measured in the baseline year only. Candidate predictors included demographic characteristics, prior healthcare utilization measures, and indicators of presence of selected medical and mental health diagnoses. Details of each of these measures are described below.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDemographic characteristics\u003c/b\u003e Demographic variables included in the models were sex and urbanicity. Sex (male vs. female) was included because of the gender differences in the course and development of medical and mental health conditions. Urbanicity was included as a binary variable to indicate whether the individual resided in a core-based statistical area (CBSA), as defined by the U.S. White House Office of Management and Budget and using population counts collected in the decennial Census.\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e In this approach, enrollees are considered as living in an urban setting if they reside in an area with a population of at least 50,000.\u003c/p\u003e \u003cp\u003e \u003cb\u003eHealthcare Utilization Measures\u003c/b\u003e Inclusion of prior healthcare utilization measures can improve model performance.\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e We included four dichotomous variables to capture prior healthcare utilization in the baseline year: all-cause hospitalization, hospitalization for a psychiatric condition, all-cause ED visit, and ED visit for a psychiatric condition.\u003c/p\u003e \u003cp\u003e \u003cb\u003eHierarchical Condition Categories (HCCs)\u003c/b\u003e We used the 70 condition categories defined in Version 12 of the CMS HCC model, which was in use for claims incurred during the study observation window (See Supplementary Material Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e HCC risk adjustment systems have been used to adjust capitation payments to Medicare Advantage Plans for the health expenditure risk for their enrollees.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e It has also been used in health services research to predict outcomes such as hospitalization risk\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e and other utilization outcomes such as emergency department use, hospital readmission, and the likelihood of being a high cost enrollee.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e For related condition categories in HCCs, hierarchies are imposed so that a patient with multiple related conditions is coded for only the most severe one. The unrelated condition categories in the HCCs are additive; an individual may be coded for none, one, or multiple HCCs.\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003e \u003cb\u003ePsyCMS psychiatric condition categories\u003c/b\u003e Many mental health and substance use conditions do not map to conditions specified in the HCC model.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e In order to more adequately capture the mental health and substance use conditions not captured in the CMS HCC V12 model, we incorporated the 46 mental health and substance use condition categories defined in the PsyCMS Case-Mix System (See Supplementary Material Table S2). This model was developed using a national sample of Department of Veterans Affairs (VA) patients with mental health and substance use disorders.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e The PsyCMS system also imposes hierarchies, based on the clinical assessment of severity, medical diagnostic criteria, and greater specificity, to reduce overlap among closely related diagnosis codes,\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e which allows us to account for both multimorbidity and the overlap among conditions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eMental Health Outcomes\u003c/b\u003e We modeled 4 binary outcomes reflecting adverse mental health outcomes in the 12- and 24-month prediction period: 1- and 2-year risk of psychiatric hospitalization, and 1- and 2-year risk of ED utilization for a psychiatric condition. Psychiatric hospitalization or ED visit were identified if the primary diagnosis for the hospitalization or ED visit was one of the psychiatric condition codes identified by the psychiatric codes used in previous studies.\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eOur predictive modeling process involved fitting a series of models, and then, for each outcome, identifying the one with the best performance as the final model. The first set of models (Models 1 through 4) were logistic regression models, with different combinations of predictors to evaluate the relative importance of each group of variables. Predictors of Model 1 were demographics (sex and urbanicity) and healthcare utilization measures. Predictors of Model 2 were demographics and HCC conditions. Predictors of Model 3 were demographics and PsyCMS conditions. Predictors of Model 4 were demographics, healthcare utilization measures, HCC conditions, and PsyCMS conditions.\u003c/p\u003e \u003cp\u003eThe second set of models (Models 5 through 7) were machine learning models using all predictors (demographics, healthcare utilization measures, HCC conditions, and PsyCMS conditions). In addition, Model 5, which utilized logistic regression with Least Absolute Shrinkage and Selection Operator (LASSO) penalty,\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e also included two-way interactions of all predictors with sex and with urbancity. Model 6 was a random forests (RF) model\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e; and Model 7 was a gradient boosting machine (GBM) model.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e We chose two classes of ML algorithms: a regression-based method (LASSO) that models additive effects of predictors along with simple two-way interactions of predictors with sex and a marker for rural/urban residence; and methods based on decision trees predictors (RF and GBM) which model complex higher order interactions involving predictors. The LASSO model is a regression-based method that, by applying shrinkage to the coefficients in the model, achieves better prediction accuracy due to reduced variance.\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e RF is an ensemble of decision trees that fits uncorrelated decision trees using bootstrapped samples and then combines them using a technique called bagging.\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e GBM is also an ensemble learning method which uses boosting to produce a complex aggregated model by combining simple prediction models.\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eWe used a 2:1 ratio for the train-test split: 2/3 of the sample was randomly selected as the training set for model development, whereas the remaining 1/3 was used as the test set to evaluate the model performance. To train the models on balanced data, the training set was downsampled by randomly selecting nonevents to match the events in a 1:1 ratio for each outcome.\u003c/p\u003e \u003cp\u003eBefore fitting each model, predictors with a frequency ratio (i.e., the ratio of the frequency of the most common category and the other category)\u0026thinsp;\u0026gt;\u0026thinsp;20 were dropped because of the low information content. All models were trained on the training set, with a five-fold cross-validation to tune the parameters for machine learning models, and model performance metrics (sensitivity, specificity, and area under the receiver operating characteristics curve (AUC)) were estimated on the test set. The sensitivity and specificity of our models were calculated by choosing a threshold for the risk prediction functions that maximized the Youden index.\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e All logistic regression models were examined for calibration (agreement between predicted and observed frequencies) using the Brier score and visual inspection, and then, if needed, re-calibrated coefficients were obtained on the training data.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eOverall, 71,682 adults aged between 55 to 64 years with a depression diagnosis and with at least 36 months of continuous enrollment met our inclusion criteria. After downsampling, for the 1-year risk of psychiatric hospitalization analysis, the training set contained 1,128 patients and the test set contained 23,893 patients. For the 2-year risk of psychiatric hospitalization analysis, the training set contained 1,904 patients and the test set contained 23,893 patients. For the 1-year risk of psychiatric ED visit analysis, the training set contained 1,306 patients and the test set contained 23,893 patients. For the 2-year risk of psychiatric ED visit analysis, the training set contained 2,282 patients and the test set contained 23,893 patients.\u003c/p\u003e \u003cp\u003eAmong the 71,682 patients included in the study cohort, 1.18% (n\u0026thinsp;=\u0026thinsp;845) experienced a psychiatric hospitalization in the 1-year prediction window, 1.99% (n\u0026thinsp;=\u0026thinsp;1,427) experienced a psychiatric hospitalization in the 2-year prediction window, 1.37% (n\u0026thinsp;=\u0026thinsp;979) utilized ED services for a psychiatric condition within the 1-year prediction window, and 2.39% (n\u0026thinsp;=\u0026thinsp;1,711) utilized ED services for a psychiatric condition in the 2-year prediction window. Approximately 70% of the sample was female, and all were aged 55\u0026ndash;64 years during the observation period. Nearly 90% resided in a metropolitan core-based statistical area.\u003c/p\u003e \u003cp\u003eThe model performance (AUC) of the seven models for each of the four outcomes are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The ROC curves of the logistic regression models for the four outcomes are presented in Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e1\u003c/span\u003ea-d, respectively.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison of Prospective Model Performance Using Area under ROC Curve (AUC) for 1- and 2-year risk of Psychiatric Hospitalization and Emergency Department Use for Psychiatric Conditions\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"7\" nameend=\"c8\" namest=\"c2\"\u003e \u003cp\u003eModel Variables used in Alternative Hospitalization Prediction Models*\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOutcome\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel 1\u0026thinsp;=\u0026thinsp;Dem. and prior utilization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eModel 2\u0026thinsp;=\u0026thinsp;Dem. and HCC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eModel 3\u0026thinsp;=\u0026thinsp;Dem. and PsyCMS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eModel 4\u0026thinsp;=\u0026thinsp;Dem., HCC and PsyCMS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eModel 5\u0026thinsp;=\u0026thinsp;\u003cb\u003eFinal model\u003c/b\u003e\u0026Dagger; utilizing machine learning \u0026ndash; logistic regression with Least Absolute Shrinkage and Selection Operator (LASSO) penalty\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eModel 6 utilizing machine learning \u0026ndash; random forests\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eModel 7 utilizing machine learning \u0026ndash; gradient boosting\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1-year risk of psychiatric hospitalization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.715\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.711\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.781\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.792\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.806\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.727\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.756\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2-year risk of psychiatric hospitalization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.694\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.694\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.757\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.777\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.781\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.715\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.749\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1-year risk of psychiatric emergency department use\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.648\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.662\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.713\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.725\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.727\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.677\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.699\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2-year risk of psychiatric emergency department use\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e.673\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e.679\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e.717\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e.737\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e.748\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e.692\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e.727\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003e*All predictors and risk data sources were derived from year 1\u003c/p\u003e \u003cp\u003e\u0026dagger; 1-year prospective outcome data was measured using claims data from year 2, and 2-year prospective outcome data was measured using claims data from years 2 and 3 combined\u003c/p\u003e \u003cp\u003e\u0026Dagger; The Final model included all variables specified in other models plus interaction terms, after applying lasso and filtering\u003c/p\u003e \u003cp\u003eDem. indicates demographics; HCC indicates Hierarchical Condition Categories; PsyCMS indicates Psychiatric Case-Mix System conditions\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFor each of the four outcomes, our best performing model was Model 5 (hereinafter referred to as the Final Model), the machine learning approach of logistic regression with LASSO penalty, using demographic characteristics, healthcare utilization measures, HCC conditions, PsyCMS conditions, and interaction effects between sex and all other variables and between urbanicty and all other variables as potential predictors.\u003c/p\u003e \u003cp\u003eFor the 1-year risk of psychiatric hospitalization, our final model achieved an AUC of 0.806, with a sensitivity of 61% and a specificity of 85% obtained under the optimal threshold 0.56. For the 2-year risk of psychiatric hospitalization, our final model achieved an AUC of 0.781, with a sensitivity 68% and a specificity of 75% under the optimal threshold 0.43. For the 1-year risk of psychiatric ED visit, our final model achieved an AUC of 0.727, with a sensitivity of 56% and a specificity of 78% under the optimal threshold 0.53. For the 2-year risk of psychiatric ED visit, our final model achieved an AUC of 0.748, with a sensitivity of 67% and a specificity of 72% using the optimal threshold of 0.47.\u003c/p\u003e \u003cp\u003eOur final models (Model 5), logistic regressions with LASSO penalty, allowed us to identify main effects and interaction effects that affected the odds of experiencing psychiatric hospitalization or ED utilization. Figures\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e2\u003c/span\u003ea-d present risk factors selected for each of the four outcomes, along with the odds ratios and 95% confidence intervals associated with these risk factors in our final models. The following factors increased the odds of experiencing all four outcomes: prior all-cause ED utilization, the PsyCMS category of anxiety not otherwise specified, the PsyCMS category of bipolar disorders, and the HCC category of major depressive and bipolar and paranoid disorders. The following factors increased the odds of three of the outcomes (i.e., 1-year psychiatric hospitalization, 2-year psychiatric hospitalization, and 2-year psychiatric ED utilization): prior psychiatric hospitalization, and the HCC category of drug/alcohol dependence. The PsyCMS category of severe Major Depressive Disorder emerged as an important risk factor in the final models for three of the four outcomes (specifically, in the final models for 1-year psychiatric hospitalization, 2-year psychiatric hospitalization, and 1-year psychiatric ED utilization). Organic Other conditions (a group of conditions specified in the PsyCMS system that include conditions such as alcohol-induced mental disorders, delusions, withdrawal and withdrawal delirium, various drug-induced mental disorders, various other organic psychotic conditions, and other nonpsychotic mental disorders due to organic brain damage) were an important risk factor in the final models for both the 1-year and 2-year psychiatric hospitalization. A diagnosis of unspecified psychosis (the condition category of Psychosis NOS in the PsyCMS system) was an important factor in the final models for both 1-year and 2-year psychiatric ED utilization. Certain mental health conditions and interactions emerged as important factors in the final models for only one of the four outcomes: the PsyCMS category of major psychiatric disorder in remission, and the interaction between Generalized Anxiety Disorder and sex were important in the 1-year psychiatric hospitalization final model; the PsyCMS category of Depression not otherwise specified, and the interaction between the PsyCMS category of Major Depressive Disorder (Single Episode) and sex was important in the 2-year psychiatric hospitalization final model; and the PsyCMS category of panic/agoraphobia emerged as an important risk factor in the 2-year psychiatric ED utilization final model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e2\u003c/span\u003ea: Selected risk factors for 1-year risk of psychiatric hospitalization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI. Risk factors above the blue dashed line correspond to the interaction between Generalized Anxiety and Sex. Marginal estimates of risk are presented for the four categories of the interaction effect adjusting for other risk factors in the model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e2\u003c/span\u003eb: Selected risk factors for 2-year risk of psychiatric hospitalization: The risk factors selected by the LASSO model (Model 5) are presented along with their Odds Ratio and 95% CI. Risk factors above the blue dashed line correspond to the interaction between Major Depressive Disorder (Single Episode) and Sex. Marginal estimates of risk are presented for the four categories of the interaction effect adjusting for other risk factors in the model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eImportantly, we were able to identify medical conditions that increased the odds of the psychiatric outcomes modeled in this paper. The HCC category of chronic obstructive pulmonary disease (COPD) emerged as a risk factor in both the 1-and 2-year final models for psychiatric hospitalization; and the HCC categories of polyneuropathy and congestive heart failure were identified as important factors in the 2-year psychiatric ED utilization final model.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed custom calibrated models to predict psychiatric hospitalization and ED utilization for psychiatric conditions among commercially insured late middle-aged adults with depression. Our models involved the use of a large set of clinical, demographic, and prior healthcare utilization variables to provide a rich description of the patients that can be used for identifying patients for care management. One benefit of our approach is that it can be easily modified to include other health conditions or demographic characteristics, such as race and household income, when available.\u003c/p\u003e \u003cp\u003eWe found that it was preferable to use a more complete method for the classification of medical and mental health and behavioral health factors, which is consistent with prior studies.\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e,\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e We found that the inclusion of the psychiatric condition categories, in addition to the HCC conditions, enhanced the performance of our models, thus allowing us to better identify high-risk patients.\u003c/p\u003e \u003cp\u003eOur best performing models for each of the four outcomes modeled (1- and 2-year risk of psychiatric hospitalization, and 1-and 2-year risk of ED utilization for psychiatric conditions) used the machine learning approach of logistic regression with LASSO penalty. This approach outperformed the other two candidate machine learning approaches of RF and GBM, indicating that a regression-based model with additive effects of predictors along with hypothesized two-way interaction effects between all predictors with sex and urbanicity can sufficiently capture the variability in the outcomes, without the need to include complex higher-order interaction terms by using RF or GBM. In addition, logistic regression with LASSO penalty produces findings that are transparent with interpretable model parameters and the approach allows for the exploration of simple interactions among model variables.\u003c/p\u003e \u003cp\u003eUnsurprisingly, several psychiatric comorbidities emerged as significant risk factors across several of the outcomes measured (e.g., severe major depressive, bipolar and paranoid disorders, anxiety, and drug/alcohol dependence, psychoses, psychoses and organic other conditions). We also found that interaction effects were observed between sex and having a single episode of Major Depressive Disorder, and between sex and Generalized Anxiety Disorder, indicating that females with these conditions are at higher risk. Further, we identified several medical comorbidities that increased the risk of psychiatric hospitalization and ED utilization, namely COPD (for the two psychiatric hospitalization outcomes), vascular disease (for 2-year psychiatric hospitalization), and polyneuropathy and congestive heart failure (for 2-year psychiatric ED utilization).\u003c/p\u003e \u003cp\u003eStrengths of our study include the use of insurance claims from multiple large U.S. insurers rather than from a single insurer or geographic region, allowing for wider applicability to commercially insured late-middle aged adults. In addition, the calculation of our risk prediction functions is easily replicable, as they rely solely on claims and demographic characteristics and do not rely on additional collection of clinical assessments or survey-based measures. Assessments and surveys may offer important prognostic variables that are not available through claims data, however availability and accessibility may depend on the standardization and reach of such assessments and the information technology environment\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e and survey-based measures often suffer from significant nonresponse bias where nonrespondents are often at higher risk of health status decline and subsequent utilization.\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThere are several limitations of the current study. First, we did not have access to certain demographic characteristics (e.g., race, ethnicity) to protect patient privacy. Second, the chronic illnesses in the baseline year were identified using diagnosis codes in claims for billing purposes; we acknowledge that, in claims data, enrollees may be miscoded because of coding errors and billing purposes. Third, our inclusion criterion of at least 36 months of continuous enrollment allowed us to predict enrollees\u0026rsquo; health outcomes over a longer time window, but might limit the generalizability of our results to the general patient population, since patients with continuous enrollment might have more stable jobs and share a different demographic and health profile. Fourth, our study focused on adults with commercial insurance, the largest payer type in the US, nevertheless, our findings may not be applicable to other payer types, such as Medicare, Medicaid, and uninsured.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eWe demonstrated that our predictive modeling approach using demographic and clinical characteristics readily available in claims data can be used to identify older adults with depression who may be at an increased risk of experiencing psychiatric hospitalization/ED utilization. Given the aging of the U.S. population and the significant burden of illness experienced by those with depression and multiple chronic conditions, this approach may assist policy makers and health care administrators in the identification of high-risk patients who may benefit from improved care delivery and monitoring.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe Institutional Review Board of Weill Cornell Medicine approved this study. The informed consent was waived by Weill Cornell Medicine Institutional Review Board because it involved secondary data analysis using deidentified data. All methods were carried out in accordance with relevant guidelines and regulations.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are available from Health Care Cost Institute (HCCI) but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of HCCI.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the National Institutes of Health (grant numbers P50MH113838, R01MH105384, T32MH073553, and K99MH130713).\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eStudy Concept and Design: SB, JP, GA, and LE\u003c/p\u003e\n\u003cp\u003eAcquisition of Data: SB and JP\u003c/p\u003e\n\u003cp\u003eAnalysis of Data: YW and MK\u003c/p\u003e\n\u003cp\u003eInterpretation of Data: LE, WX, and AG\u003c/p\u003e\n\u003cp\u003eDrafting and Revision of the Manuscript: LE, WX, and AG\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\n"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKaton WJ, Lin E, Russo J, Unutzer J. Increased medical costs of a population-based sample of depressed elderly patients. \u003cem\u003eArch Gen Psychiatry. \u003c/em\u003e2003;60(9):897-903.\u003c/li\u003e\n\u003cli\u003eUnutzer J, Patrick DL, Simon G, et al. Depressive symptoms and the cost of health services in HMO patients aged 65 years and older. A 4-year prospective study. \u003cem\u003eJama. \u003c/em\u003e1997;277(20):1618-1623.\u003c/li\u003e\n\u003cli\u003eHimelhoch S, Weller WE, Wu AW, Anderson GF, Cooper LA. Chronic medical illness, depression, and use of acute medical services among Medicare beneficiaries. \u003cem\u003eMed Care. \u003c/em\u003e2004;42(6):512-521.\u003c/li\u003e\n\u003cli\u003eHuang BY, Cornoni-Huntley J, Hays JC, Huntley RR, Galanos AN, Blazer DG. Impact of depressive symptoms on hospitalization risk in community-dwelling older persons. \u003cem\u003eJ Am Geriatr Soc. \u003c/em\u003e2000;48(10):1279-1284.\u003c/li\u003e\n\u003cli\u003eEttner SL. The setting of psychiatric care for medicare recipients in general hospitals with specialty units. \u003cem\u003ePsychiatric services (Washington, DC). \u003c/em\u003e2001;52(2):237-239.\u003c/li\u003e\n\u003cli\u003ePrince JD, Akincigil A, Kalay E, et al. Psychiatric rehospitalization among elderly persons in the United States. \u003cem\u003ePsychiatric services (Washington, DC). \u003c/em\u003e2008;59(9):1038-1045.\u003c/li\u003e\n\u003cli\u003e\u003cem\u003eHeslin KC (AHRQ), Weiss AJ (Truven Health Analytics). Hospital Readmissions Involving Psychiatric Disorders, 2012. HCUP Statistical Brief #189. May 2015. Agency for Healthcare Research and Quality, Rockville, MD. \u003c/em\u003e\u003cem\u003ehttp://www.hcup-us.ahrq.gov/reports/statbriefs/sb189-Hospital-Readmissions-Psychiatric-Disorders-2012.pdf\u003c/em\u003e\u003cem\u003e.\u003c/em\u003e\u003c/li\u003e\n\u003cli\u003eWeiss AJ, Barrett ML, Heslin KC, Stocks C. \u003cem\u003eTrends in Emergency Department Visits Involving Mental and Substance Use Disorders, 2006\u0026ndash;2013.HCUP Statistical Brief #216. December 2016. Agency for Healthcare Research and Quality, Rockville, MD. Accessed December 23, 2019. Available at: \u003c/em\u003e\u003cem\u003ehttp://www.hcup-us.ahrq.gov/reports/statbriefs/sb216-Mental-Substance-Use-Disorder-ED-Visit-Trends.pdf\u003c/em\u003e\u003cem\u003e. .\u003c/em\u003e\u003c/li\u003e\n\u003cli\u003eOwens PL, Heslin KC, Fingar KR, Weiss AJ. \u003cem\u003eCo-occurrence of Physical Health Conditions and Mental Health and Substance Use Conditions Among Adult Inpatient Stays, 2010 Versus 2014. HCUP Statistical Brief #240. June 2018. Agency for Healthcare Research and Quality, Rockville, MD. \u003c/em\u003e\u003cem\u003ewww.hcup-us.ahrq.gov/reports/statbriefs/sb240-Co-occurring-Physical-Mental-Substance-Conditions-Hospital-Stays.pdf\u003c/em\u003e\u003cem\u003e.\u003c/em\u003e\u003c/li\u003e\n\u003cli\u003eStrom JB, Kramer DB, Wang Y, et al. Short-term rehospitalization across the spectrum of age and insurance types in the United States. \u003cem\u003ePLoS One. \u003c/em\u003e2017;12(7):e0180767.\u003c/li\u003e\n\u003cli\u003eTsai J, Rosenheck RA, Culhane DP, Artiga S. Medicaid expansion: chronically homeless adults will need targeted enrollment and access to a broad range of services. \u003cem\u003eHealth Aff (Millwood). \u003c/em\u003e2013;32(9):1552-1559.\u003c/li\u003e\n\u003cli\u003eOwens PL, Fingar KR, McDermott KW, Muhuri PK, Heslin KC. Inpatient Stays Involving Mental and Substance Use Disorders, 2016: Statistical Brief #249. In: \u003cem\u003eHealthcare Cost and Utilization Project (HCUP) Statistical Briefs.\u003c/em\u003e Rockville (MD): Agency for Healthcare Research and Quality (US); 2006.\u003c/li\u003e\n\u003cli\u003ePerlman CM, Hirdes JP, Vigod S. Psychiatric Rehospitalization: Development of a Person-Level Indicator for Care Planning and Quality Assurance. \u003cem\u003ePrim Care Companion CNS Disord. \u003c/em\u003e2015;17(4).\u003c/li\u003e\n\u003cli\u003ePope GC, Kautter J, Ellis RP, et al. Risk adjustment of Medicare capitation payments using the CMS-HCC model. \u003cem\u003eHealth care financing review. \u003c/em\u003e2004;25(4):119-141.\u003c/li\u003e\n\u003cli\u003eSloan KL, Montez-Rath ME, Spiro A, 3rd, et al. Development and validation of a psychiatric case-mix system. \u003cem\u003eMed Care. \u003c/em\u003e2006;44(6):568-580.\u003c/li\u003e\n\u003cli\u003eNewman D, Herrera CN, Parente ST. Overcoming barriers to a research-ready national commercial claims database. \u003cem\u003eAm J Manag Care. \u003c/em\u003e2014;20(11 Spec No. 17):eSP25-30.\u003c/li\u003e\n\u003cli\u003eTownsend L, Walkup JT, Crystal S, Olfson M. A systematic review of validated methods for identifying depression using administrative data. \u003cem\u003ePharmacoepidemiol Drug Saf. \u003c/em\u003e2012;21 Suppl 1:163-173.\u003c/li\u003e\n\u003cli\u003eFiest KM, Jette N, Quan H, et al. Systematic review and assessment of validated case definitions for depression in administrative data. \u003cem\u003eBMC Psychiatry. \u003c/em\u003e2014;14:289.\u003c/li\u003e\n\u003cli\u003eHall SA, Kaufman JS, Ricketts TC. Defining urban and rural areas in U.S. epidemiologic studies. \u003cem\u003eJ Urban Health. \u003c/em\u003e2006;83(2):162-175.\u003c/li\u003e\n\u003cli\u003eEverson J, Hollingsworth JM, Adler-Milstein J. Comparing methods of grouping hospitals. \u003cem\u003eHealth Serv Res. \u003c/em\u003e2019;54(5):1090-1098.\u003c/li\u003e\n\u003cli\u003eLemke KW, Weiner JP, Clark JM. Development and validation of a model for predicting inpatient hospitalization. \u003cem\u003eMed Care. \u003c/em\u003e2012;50(2):131-139.\u003c/li\u003e\n\u003cli\u003e\u003cem\u003eEvaluation of the CMS-HCC Risk Adjustment Model: Final Report. Authors: Pope GC, Kautter J, Ingber MJ, Freeman S, Sekar R, Newhart C. Federal Project Officer: Melissa A. Evans, PhD. RTI International. CMS Contract No. HHSM-500-2005-00029I TO 0006. March 2011.\u003c/em\u003e\u003c/li\u003e\n\u003cli\u003eRosen AK, Loveland SA, Anderson JJ, Hankin CS, Breckenridge JN, Berlowitz DR. Diagnostic cost groups (DCGs) and concurrent utilization among patients with substance abuse disorders. \u003cem\u003eHealth services research. \u003c/em\u003e2002;37(4):1079-1103.\u003c/li\u003e\n\u003cli\u003eMosley DG, Peterson E, Martin DC. Do hierarchical condition category model scores predict hospitalization risk in newly enrolled Medicare advantage participants as well as probability of repeated admission scores? \u003cem\u003eJ Am Geriatr Soc. \u003c/em\u003e2009;57(12):2306-2310.\u003c/li\u003e\n\u003cli\u003eHaas LR, Takahashi PY, Shah ND, et al. Risk-stratification methods for identifying patients for care coordination. \u003cem\u003eAm J Manag Care. \u003c/em\u003e2013;19(9):725-732.\u003c/li\u003e\n\u003cli\u003eWagner TH, Almenoff P, Francis J, Jacobs J, Pal Chee C. Assessment of the Medicare Advantage Risk Adjustment Model for Measuring Veterans Affairs Hospital Performance. \u003cem\u003eJAMA Netw Open. \u003c/em\u003e2018;1(8):e185993.\u003c/li\u003e\n\u003cli\u003eKautter J, Pope GC, Ingber M, et al. The HHS-HCC risk adjustment model for individual and small group markets under the Affordable Care Act. \u003cem\u003eMedicare Medicaid Res Rev. \u003c/em\u003e2014;4(3).\u003c/li\u003e\n\u003cli\u003eJohnston KJ, Allen L, Melanson TA, Pitts SR. A \u0026quot;Patch\u0026quot; to the NYU Emergency Department Visit Algorithm. \u003cem\u003eHealth Serv Res. \u003c/em\u003e2017;52(4):1264-1276.\u003c/li\u003e\n\u003cli\u003eHastie T, Tibshirani R, Friedman J. Linear Methods for Regression. In: \u003cem\u003eThe Elements of Statistical Learning: Data Mining, Inference, and Prediction.\u003c/em\u003e 2nd ed. New York: Springer; 2009.\u003c/li\u003e\n\u003cli\u003eHastie T, Tibshirani R, Friedman J. Random Forests. In: \u003cem\u003eThe Elements of Statistical Learning: Data Mining, Inference, and Prediction.\u003c/em\u003e 2nd ed. New York: Springer; 2009.\u003c/li\u003e\n\u003cli\u003eHastie T, Tibshirani R, Friedman J. Boosting and Additive Trees. In: \u003cem\u003eThe Elements of Statistical Learning: Data Mining, Inference, and Prediction.\u003c/em\u003e 2nd ed. New York: Springer; 2009.\u003c/li\u003e\n\u003cli\u003eR T. Regression shrinkage and selection via the Lasso \u003cem\u003eJournal of the Royal Statistical Society. \u003c/em\u003e1996;58(1):267-288.\u003c/li\u003e\n\u003cli\u003eFriedman JH. Greedy function approximation: A gradient boosting machine. \u003cem\u003eAnn Statist. \u003c/em\u003e2001;29(5):1189-1232.\u003c/li\u003e\n\u003cli\u003eBreiman L. Random Forests. \u003cem\u003eMachine Learning. \u003c/em\u003e2001;45(1):5-32.\u003c/li\u003e\n\u003cli\u003eFluss R, Faraggi D, Reiser B. Estimation of the Youden Index and its associated cutoff point. \u003cem\u003eBiom J. \u003c/em\u003e2005;47(4):458-472.\u003c/li\u003e\n\u003cli\u003eShrestha A, Bergquist S, Montz E, Rose S. Mental Health Risk Adjustment with Clinical Categories and Machine Learning. \u003cem\u003eHealth Serv Res. \u003c/em\u003e2018;53 Suppl 1:3189-3206.\u003c/li\u003e\n\u003cli\u003eEttner SL, Frank RG, McGuire TG, Hermann RC. Risk adjustment alternatives in paying for behavioral health care under Medicaid. \u003cem\u003eHealth services research. \u003c/em\u003e2001;36(4):793-811.\u003c/li\u003e\n\u003cli\u003eKinosian B, Wieland D, Gu X, Stallard E, Phibbs CS, Intrator O. Validation of the JEN frailty index in the National Long-Term Care Survey community population: identifying functionally impaired older adults from claims data. \u003cem\u003eBMC Health Serv Res. \u003c/em\u003e2018;18(1):908.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"insurance claims data, psychiatric hospitalization, psychiatric emergency department, service utilization, risk adjustment, late middle-aged adults, depression, machine learning","lastPublishedDoi":"10.21203/rs.3.rs-4902124/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4902124/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Late middle-aged adults with depression experience a high illness burden often resulting from multiple chronic conditions. Risk stratification using administrative claims data is a promising method to identify enrollees at highest risk of experiencing psychiatric hospitalization and emergency department (ED) utilization. The goal of this study was to develop and validate a non-proprietary model to identify late middle-aged adults with depression at risk of 1- and 2- year psychiatric hospitalization and ED utilization, with broad applicability to commercially insured adults.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e: We conducted a retrospective cohort study on 71,682 commercially insured adults aged 55-64 years with depression diagnosis. National health insurance claims were used to capture demographics, health care utilization, and health status during the baseline year. Health status was captured using 70 chronic health conditions, and 46 mental health conditions. The four outcomes were one-year and two-year psychiatric hospitalization and ED utilization.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: For 1-year risk of psychiatric hospitalization, the final model achieved an AUC of 0.806, with a sensitivity of ­­61% and a specificity of 85%. For 2-year risk of psychiatric hospitalization, the final model achieved an AUC of 0.781, with a sensitivity of 68% and a specificity of 75%. For 1-year risk of psychiatric ED utilization, the final model achieved an AUC of 0.727, with a sensitivity of 56% and specificity of 78%. For the 2-year risk of psychiatric ED utilization, the final model achieved an AUC of 0.748, with a sensitivity of 67% and a specificity of 72%. The final models for all four outcomes outperformed their corresponding models using only demographics and prior utilization.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003ePredictive modeling using diagnoses and other demographic characteristics readily available in claims data can be used to identify late middle-aged adults with depression at risk of psychiatric hospitalization and ED utilization. Given the aging of the U.S. population and the significant burden of illness experienced by those with depression and multiple chronic conditions, this approach may assist policy makers and health care administrators in the identification of high-risk patients who may benefit from additional screening and care management interventions.\u003c/p\u003e","manuscriptTitle":"Predictions of psychiatric hospitalization and emergency department utilization in commercially insured late middle-aged adults with depression","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-10-08 02:34:19","doi":"10.21203/rs.3.rs-4902124/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"aa811eb2-e0ce-4e25-87bc-30de8c9229a8","owner":[],"postedDate":"October 8th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-10-06T08:09:17+00:00","versionOfRecord":[],"versionCreatedAt":"2024-10-08 02:34:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4902124","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4902124","identity":"rs-4902124","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}