Abstract
Objectives
Patients with substance misuse are at high risk for clinical deterioration, and pre-hospital encounters constitute important risk factors. We sought to incorporate these risk factors into a novel prediction model using linked electronic health record, emergency medical services (EMSs), and claims data.
Materials and methods
Using 23 454 hospital encounters, we developed machine-learning models to predict mechanical ventilation, vasopressor initiation, or death within 12 hours of a vital sign or laboratory measurement. Models were compared to the modified early warning score.
Results
Extreme gradient boosting achieved an area under the receiver operating characteristic curve of 0.93, significantly outperforming modified early warning score (MEWS 0.79). Notably, removing EMS and claims data did not reduce predictive performance.
Discussion
The model demonstrated strong predictive performance of clinical deterioration in this high-risk population. Incorporation of pre-hospital risk factors did not improve performance beyond EHR data alone.
Conclusion
Electronic health record–based machine learning models can support accurate short-term prediction of clinical deterioration among this population at risk for delayed recognition.
Keywords: prediction methods, machine, medical record linkage, electronic health records, critical illness, emergency medical services
Introduction
Emergency department visits and hospitalizations involving patients with substance misuse have increased steadily in recent years, reflecting a growing demand for high-acuity care for these patients.1 Despite being younger than the general inpatient population, these patients experience higher rates of critical illness and in-hospital mortality.2 Stigma surrounding this population may contribute to these disparities, as studies have shown that these patients experience poorer communication, delayed or withheld treatment, and assumptions that their symptoms reflect behavioral issues rather than medical needs.3,4 Notably, prior work has shown that stigmatizing language in the clinical notes of septic patients with substance misuse is associated with higher mortality, illustrating how stigma might lead to suboptimal care.5 These findings underscore the need for reliable, objective systems to recognize early clinical decline in this population.
In addition to challenges within the hospital setting, risk factors for poor outcomes among patients with substance misuse may also be influenced by factors occurring before admission. Prior studies have shown that recent emergency medical service (EMS) utilization, emergency department (ED) visits, and hospitalizations are independent predictors of adverse in-hospital outcomes in this population. This suggests that clinical deterioration risk may accumulate across the care continuum rather than begin at hospital admission.6,7 Including these pre-admission risk factors was shown to improve a model forecasting 30-day readmission and mortality at discharge among patients with substance misuse.8 However, no study has examined whether adding these factors to risk prediction models improves early recognition of in-hospital deterioration. Accordingly, this study sought to integrate pre-hospital data sources, such as EMS data and inpatient insurance claims, with electronic health records (EHRs), to develop a machine learning model predicting critical illness or death among hospitalized patients with substance misuse and compare its performance to that of an existing early warning system. The secondary objective was to evaluate the added value of pre-hospital data for predicting inpatient deterioration.
Methods
Study population
We conducted a retrospective cohort study of hospital encounters in the Substance Misuse Data Commons (SMDC).9 The SMDC includes the ED visits and inpatient hospitalizations at 2 University of Wisconsin–Madison hospitals, which are regional tertiary care academic medical centers serving a catchment of urban and rural populations, between 2008 and 2022 for all adult (>18 years of age) patients with at least one substance-misuse-related International Classification of Diseases diagnosis code (see Table S1).10 The repository links EHR data with data from state agencies, medical and pharmacy claims, neighborhood socioeconomic indicators, prescription drug monitoring systems, and national mortality data. We examined data from 2017 to 2021 because this spanned the years with complete EHR, claims, and EMS data. Patients were excluded if they did not have any vital signs or laboratory data during the hospital encounter. This study was approved by the University of Wisconsin-Madison Internal Review Board (IRB 2021-0553), which did not require explicit consent from the participants. All methods were performed in accordance with the relevant guidelines and regulations. Reporting criteria followed Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence (TRIPOD + AI) for abstracts.11 The study was not prospectively registered in a public registry.
Data collection
We extracted structured EHR, EMS, and medical claims data from the SMDC, and all variables underwent standard data quality checks before analysis. These variables included patient demographics, hospital arrival mode, health insurance provider, hospitalization in the past 30 days, national area deprivation index (ADI, a neighborhood socioeconomic measure derived from patients’ home addresses), Elixhauser comorbidities, preceding EMS incidents, vital signs, laboratory results, and hospital location (eg, ward, ED or intensive care unit).12,13 Prehospital vital signs were carried forward to impute missing initial values. Hospital vital signs and laboratory results were carried forward over time, and the remaining missing values were imputed using the hospital location-specific medians estimated from the training data. Missing ADI values were imputed from the training set median. Training data were then blocked into 12-hour intervals following a discrete-time survival analysis framework used in prior studies.14
For comparisons of variables in Table 1 and Supplementary Table 3, P-values were calculated using the Kruskal–Wallis rank-sum test for continuous variables and Pearson’s chi-square test for categorical variables.
Table 1.
Cohort characteristics.
| Encounters without critical illness or death | Encounters with critical illness or death | Total encounters | P-value | |
|---|---|---|---|---|
| Age median (IQR) | 46 (32, 59) | 53 (39, 62) | 47 (33, 59) | <.001 |
| Male, n (%) | 12 788 (62%) | 1777 (67%) | 14 565 (63%) | <.001 |
| Race, n (%) | <.001 | |||
| American Indian or Alaska Native | 328 (1.6%) | 46 (1.7%) | 374 (1.6%) | |
| Asian/Mideast Indian | 228 (1.1%) | 18 (0.7%) | 246 (1.1%) | |
| Black/African American | 2476 (12%) | 197 (7.4%) | 2673 (11%) | |
| Other | 238 (1.2%) | 74 (2.8%) | 312 (1.3%) | |
| Pacific Islander/Hawaiian Native | 47 (0.2%) | 2 (<0.1%) | 49 (0.2%) | |
| White/Caucasian | 17 316 (84%) | 2328 (87%) | 19 644 (84%) | |
| Means of arrival, n (%) | <.001 | |||
| EMS | 7726 (37%) | 1229 (46%) | 8955 (38%) | |
| Police | 845 (4.1%) | 28 (1.1%) | 873 (3.7%) | |
| Private transport | 7356 (36%) | 269 (10%) | 7625 (33%) | |
| Other | 4706 (23%) | 1139 (43%) | 5845 (25%) | |
| Payor, n (%) | <.001 | |||
| Medicare | 4835 (23%) | 552 (21%) | 5387 (23%) | |
| Medicaid | 5782 (28%) | 542 (20%) | 6324 (27%) | |
| Commercial | 5983 (29%) | 599 (22%) | 6582 (28%) | |
| Other | 146 (0.7%) | 32 (1.2%) | 178 (0.8%) | |
| NA | 3887 (19%) | 940 (35%) | 4827 (21%) | |
| ADI National Rank Median (IQR) | 46 (32, 59) | 49 (39, 65) | 46 (32, 59) | <.001 |
| Hospitalization in the past 30 days, n (%) | 5566 (27%) | 652 (24%) | 6218 (27%) | .006 |
| Comorbidities at, n (%) | <.001 | |||
| Congestive heart failure | 2562 (12%) | 448 (17%) | 3010 (13%) | |
| Cardiac arrythmias | 10 099 (49%) | 1105 (41%) | 11 204 (48%) | |
| Valvular disease | 2882 (14%) | 441 (17%) | 3323 (14%) | |
| Pulmonary circulation disorders | 2482 (12%) | 365 (14%) | 2847 (12%) | |
| Peripheral vascular disorders | 3175 (15%) | 500 (19%) | 3675 (16%) | |
| Hypertension, uncomplicated | 9399 (46%) | 1240 (47%) | 10 639 (46%) | |
| Hypertension, complicated | 2671 (13%) | 402 (15%) | 3073 (13%) | |
| Paralysis | 673 (3.3%) | 76 (2.9%) | 749 (3.2%) | |
| Other neurologic disorders | 4958 (24%) | 678 (25%) | 5636 (24%) | |
| Chronic pulmonary disease | 7492 (36%) | 915 (34%) | 8407 (36%) | |
| Diabetes, uncomplicated | 3239 (16%) | 499 (19%) | 3738 (16%) | |
| Diabetes, complicated | 2319 (11%) | 345 (13%) | 2664 (11%) | |
| Hypothyroidism | 2284 (11%) | 274 (10%) | 2558 (11%) | |
| Renal failure | 2493 (12%) | 427 (16%) | 2920 (13%) | |
| Liver disease | 7061 (34%) | 958 (36%) | 8019 (34%) | |
| Peptic ulcer disease | 1618 (7.8%) | 217 (8.1%) | 1835 (7.9%) | |
| AIDS/HIV | 313 (1.5%) | 14 (0.5%) | 327 (1.4%) | |
| Lymphoma | 334 (1.6%) | 47 (1.8%) | 381 (1.6%) | |
| Metastatic cancer | 1045 (5.1%) | 137 (5.1%) | 1182 (5.1%) | |
| Solid tumor, without metastasis | 2388 (12%) | 323 (12%) | 2711 (12%) | |
| Rheumatologic disease | 1746 (8.5%) | 195 (7.3%) | 1941 (8.3%) | |
| Coagulopathy | 3747 (18%) | 582 (22%) | 4329 (19%) | |
| Obesity | 4790 (23%) | 572 (21%) | 5362 (23%) | |
| Weight loss | 4871 (24%) | 607 (23%) | 5478 (24%) | |
| Fluid and electrolyte disorders | 9361 (45%) | 1107 (42%) | 10 468 (45%) | |
| Blood loss anemia | 1136 (5.5%) | 175 (6.6%) | 1311 (5.6%) | |
| Deficiency anemia | 3177 (15%) | 438 (16%) | 3615 (16%) | |
| Alcohol abuse | 10 839 (53%) | 1203 (45%) | 12 042 (52%) | |
| Drug abuse | 8712 (42%) | 729 (27%) | 9441 (41%) | |
| Psychoses | 2654 (13%) | 175 (6.6%) | 2829 (12%) | |
| Depression | 11 929 (58%) | 1110 (42%) | 13 039 (56%) | |
| EMS incidents in the past 30 days median (IQR) | 0 (0, 0) | 0 (0, 0) | 0 (0, 0) | .14 |
| EMS incident in the past 24 hours, n (%) | 8338 (40%) | 1714 (64%) | 10 052 (43%) | <.001 |
| Primary hospital location, n (%) | <.001 | |||
| Ward | 12 296 (60%) | 1711 (64%) | 14 007 (60%) | |
| Intensive care unit | 196 (0.9%) | 643 (24%) | 839 (3.6%) | |
| Emergency department | 8141 (39%) | 311 (12%) | 8452 (36%) | |
| Positive urine drug screen, n (%) | <.001 | |||
| Amphetamine | 506 (2.5%) | 73 (2.7%) | 579 (2.5%) | |
| Benzodiazepine | 675 (3.3%) | 221 (8.3%) | 896 (3.8%) | |
| Cocaine | 813 (3.9%) | 99 (3.7%) | 912 (3.9%) | |
| Opiate | 755 (3.7%) | 159 (6.0%) | 914 (3.9%) | |
| Blood alcohol tested, n (%) | 4905 (24%) | 593 (22%) | 5498 (24%) | .082 |
| Vasopressors, n (%) | 0 (0%) | 1524 (57%) | 1524 (6.5%) | — |
| Invasive mechanical ventilation, n (%) | 0 (0%) | 1942 (73%) | 1942 (8.3%) | — |
| Death, n (%) | 0 (0%) | 328 (12%) | 328 (1.4%) | — |
Abbreviations: ADI, area deprivation index; EMS, emergency medical service; IQR, interquartile range; n, numbers.
Study outcome
The primary outcome was clinical deterioration, defined as the initiation of invasive mechanical ventilation, vasopressors, or in-hospital death within 12 hours of a vital sign or laboratory measurement. Deterioration events in the operating room were excluded. Vasopressor administration was censored for 24 hours after last use, and mechanical ventilation was censored for 48 hours after last use, to exclude reintubations after failed extubations. An encounter could contribute multiple outcome events if mechanical ventilation or vasopressors were re-initiated after being off support for these washout periods.
Machine learning model development
For the primary model development, we trained logistic regression and extreme gradient boosting (XGBoost) models using all available extracted variables to predict the binary outcome. Encounters were partitioned randomly into 70% for model development and 30% for hold-out testing, stratified by outcome. Hyperparameter optimization was conducted using 5-fold cross-validation to optimize area under the receiver operating characteristic curve (AUROC), and final performance was assessed on the held-out test set. XGBoost hyperparameters (number of trees, tree depth, learning rate, subsampling, and column sampling) were tuned over a predefined grid, while logistic regression contained no tunable hyperparameters and was fit once using the same cross-validation procedure. Class probabilities were estimated for all models for ROC-based evaluation.
To assess the contribution of multi-source data, we performed an ablation study by retraining an XGBoost model after removing all EMS and claims data, using the same cross-validation and tuning strategy. To evaluate performance at initial hospital presentation, we repeated this approach using only the first available observation for each encounter. To examine model performance across hospital units, we also trained location-specific XGBoost models using the same modeling approach.
Model evaluation
Model performance was evaluated in the test set using the AUROC and the area under the precision recall curve (AUPRC). The primary multisource models were benchmarked against the modified early warning score (MEWS), using DeLong’s test.15,16 For the location-stratified analysis, we evaluated both the pooled XGBoost model and the corresponding location-specific XGBoost model within each hospital location subgroup. For the analysis of performance at initial hospital presentation, we evaluated only the first available observation from each encounter in the test set. To assess face validity, we examined feature importance in the primary multi-source XGBoost model, where importance reflects the contribution of each predictor to the model’s ROC performance. We additionally evaluated SHapley Additive exPlanations (SHAP) values for the model’s most important features to interpret the magnitude and directionality of the predictors.
Results
We analyzed 23 454 hospital encounters from the SMDC, of which 2665 (11.3%) experienced the primary outcome of clinical deterioration, defined as the initiation of invasive mechanical ventilation, vasopressors, or death within 12 hours of a vital sign or laboratory result (see Table 1). The median age of the cohort was 47 years. Overall, 38% arrived by EMS, 27% had been hospitalized in the prior 30 days, and 21% had no insurance recorded. The median national ADI was 46 (interquartile range 32–59). The most common Elixhauser comorbidities were depression (56%), alcohol use disorder (52%), and cardiac arrythmia (48%). Alcohol misuse was the most common type of substance misuse in the cohort, followed by opioid misuse (see Table S2). A total of 3919 (16.7%) hospital encounters were preceded by an EMS incident within the prior 30 days that was not directly related to the hospitalization. Most hospital encounters primarily occurred on the ward (60%), with 36% in the ED and 3.7% in the intensive care unit (ICU).
Applying the MEWS to the test cohort yielded an AUROC of 0.79 for predicting clinical deterioration (see Table 2). In comparison, the logistic regression and XGBoost models achieved AUROCs of 0.90 and 0.93, respectively. Table S3 summarizes all included predictors and their missingness. XGBoost significantly outperformed both logistic regression and MEWS (P < .001 for both comparisons). Area under the precision-recall curve (AUPRC) was the highest for the XGBoost model (see Table 2). The most influential predictors (see Figure 1) for the XGBoost model were level of consciousness, fraction of inspired oxygen, and location in the hospital. SHAP analysis demonstrated that worse level of consciousness, higher fraction of inspired oxygen, and intensive care unit location generally contributed toward higher predicted risk (see Figure S1). The location-specific models had similar AUROCs to the pooled XGBoost model within each hospital location subgroup (see Table S4). The ablation study demonstrated that removing EMS and claims data and then re-training the model did not change the AUROC or AUPRC. Models trained only on the first available observation per encounter performed similarly, and inclusion of EMS and claims data did not meaningfully improve initial-presentation performance (see Table S5).
Table 2.
Model performance on test set.
| AUROC (95% CI) | AUPRC | Sensitivity at Youden’s statistic | Specificity at Youden’s statistic | |
|---|---|---|---|---|
| MEWS | 0.79 (0.79–0.79) | 0.18 | 0.68 | 0.79 |
| EHR, EMS, claims logistic regression | 0.90 (0.89–0.90) | 0.43 | 0.79 | 0.85 |
| EHR, EMS, claims XGBoost | 0.93 (0.93–0.93) | 0.60 | 0.86 | 0.86 |
Abbreviations: AUPRC, area under the precision–recall curve; AUROC, area under the receiver operating curve; CI, confidence interval; EHR, electronic health record; EMS, emergency medical services; MEWS, modified early warning score; XGBoost, extreme gradient boosting.
Figure 1.
Variable importance of features in the multi-source XGBoost model. Variable importance for the XGBoost model predicting mechanical ventilation, vasopressor initiation, or death within 12 hours. Importance values reflect the relative contribution of each predictor to the model’s overall performance.
Discussion
This study developed a machine learning model that accurately predicts, throughout the hospital stay, the initiation of mechanical ventilation, vasopressors, or death in the next 12 hours among patients with substance misuse. The model demonstrated excellent discrimination (AUROC 0.93), outperforming both a traditional logistic regression model and the widely used MEWS. Notably, removing EMS and claims data did not affect the model’s performance, indicating that EHR data alone may be sufficient for accurate short-term inpatient predictions.
Substance misuse can alter physiologic parameters in ways that may decrease the performance of standard early warning scores. For example, intoxication or withdrawal may produce abnormal vital signs, sedative substances can mask instability, and chronic alcohol use can lead to autonomic dysfunction, factors that can make models developed on general population vital signs less specific.17 Importantly, these same physiologic patterns are also at risk of being dismissed in clinical practice when abnormal signs and symptoms are attributed to substance use, malingering, or agitation, rather than potential critical illness.3,4 This diagnostic overshadowing likely contributes to the poor clinical outcomes seen in practice, underscoring the need for deterioration models specifically trained for patients with substance misuse.
Interestingly, including EMS data and prior hospitalizations did not enhance predictive performance for this outcome. This may reflect the low prevalence of such data in the cohort; in the preceding 30 days, fewer than 17% had an EMS encounter and just over one-quarter had been hospitalized. Moreover, EHR data may already provide sufficiently granular and specific information for predicting imminent inpatient outcomes. Consistent with this interpretation, the variable importance analysis revealed that the top predictors all originated from the EHR. Although hospital location was among the most influential predictors, location-stratified analyses showed similar performance between the pooled and location-specific models, suggesting that the model may not be driven solely by differences in baseline acuity across care settings.
Limitations of this study include its development using single-center data from a population with limited racial and ethnic diversity, which may limit generalizability. Future work should include external validation in more diverse populations and in regions with different patterns of substance misuse. Additionally, the complexity of ICD-based cohort identification may limit capture of clinically meaningful variation in substance use. A strength of the study is that unlike the MEWS, our model defined critical illness by the initiation of invasive mechanical ventilation or vasopressors, the timing of which is less likely than ICU transfer to vary according to institutional practices and capacity. Additionally, as we found that EHR-only data performed well, future work should test our model against a machine learning model trained on EHR data from a general inpatient population to elucidate if patterns of deterioration differ between the groups.
Conclusion
Overall, this prediction tool leverages readily available clinical data to accurately identify substance misuse patients at high risk of deterioration. As high performance was maintained with EHR data alone, this tool may offer a scalable approach to supporting efficient resource allocation, accelerating the detection of critical illness, and improving clinical outcomes for this vulnerable patient population.
Supplementary Material
Acknowledgments
We thank our collaborators in the State of Wisconsin Department of Health Services Office of Health Informatics and the University of Wisconsin-Madison Clinical and Health Informatics Institute for contributing data to the Substance Misuse Data Commons.
Contributor Information
Preeti Gupta, Scripps Translational Research Institute, The Scripps Research Institute, La Jolla, CA 92037, United States.
Tim Gruenloh, Department of Biostatistics and Medical Informatics, University of Wisconsin-Madison, Madison, WI 53726, United States.
Madeline Oguss, School of Medicine and Public Health, Department of Medicine, University of Wisconsin-Madison, Madison, WI 53792, United States.
Askar Safipour Afshar, School of Medicine and Public Health, Department of Medicine, University of Wisconsin-Madison, Madison, WI 53792, United States.
Michael Spigner, BerbeeWalsh Department of Emergency Medicine, University of Wisconsin-Madison, Madison, WI 53792, United States.
Megan Gussick, BerbeeWalsh Department of Emergency Medicine, University of Wisconsin-Madison, Madison, WI 53792, United States.
Matthew Churpek, Department of Biostatistics and Medical Informatics, University of Wisconsin-Madison, Madison, WI 53726, United States; School of Medicine and Public Health, Department of Medicine, University of Wisconsin-Madison, Madison, WI 53792, United States.
Todd Lee, College of Pharmacy, Department of Pharmacy Systems, Outcomes, and Policy, University of Illinois Chicago, Chicago, IL 60612, United States.
Majid Afshar, Department of Biostatistics and Medical Informatics, University of Wisconsin-Madison, Madison, WI 53726, United States; School of Medicine and Public Health, Department of Medicine, University of Wisconsin-Madison, Madison, WI 53792, United States.
Anoop Mayampurath, Department of Biostatistics and Medical Informatics, University of Wisconsin-Madison, Madison, WI 53726, United States; School of Medicine and Public Health, Department of Medicine, University of Wisconsin-Madison, Madison, WI 53792, United States.
Author contributions
Preeti Gupta (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Writing—original draft, Writing—review & editing), Tim Gruenloh (Data curation, Validation, Writing—review & editing), Madeline Oguss (Investigation, Project administration, Writing—review & editing), Askar Safipour Afshar (Data curation, Writing—review & editing), Michael Spigner (Investigation, Writing—review & editing), Megan Gussick (Investigation, Writing—review & editing), Matthew Churpek (Methodology, Writing—review & editing), Todd Lee (Supervision, Writing—review & editing), Majid Afshar (Conceptualization, Funding acquisition, Investigation, Methodology, Resources, Supervision, Writing—review & editing), and Anoop Mayampurath (Conceptualization, Investigation, Methodology, Supervision, Writing—review & editing)
Supplementary material
Supplementary material is available at JAMIA Open online.
Funding
P.G. is supported by a K12 award from NCATS (K12TR004410). This project and M.A., T.G., and M.O. are supported by a R01 and supplement from NIDA (R01DA051464-S1). A.M. is supported by a R01 award from the National Heart, Lung, and Blood Institute (R01HL173037).
Conflicts of interest
None declared.
Data availability
Data availability is subject to existing data use agreements. For further details, please contact the corresponding author.
Code availability
Code is available upon reasonable request.
References
- 1. Suen LW, Makam AN, Snyder HR, et al. National prevalence of alcohol and other substance use disorders among emergency department visits and hospitalizations: NHAMCS 2014–2018. J Gen Intern Med. 2022;37:2420-2428. 10.1007/s11606-021-07069-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Steel TL, Bhatraju EP, Hills-Dunlap K. Critical care for patients with substance use disorders. Curr Opin Crit Care. 2023;29:484-492. 10.1097/MCC.0000000000001080 [DOI] [PubMed] [Google Scholar]
- 3. Rajab D, Fujioka JK, Walker M, Bartels SA, MacKenzie M, Purkey E. Emergency department care experiences among people who use substances: a qualitative study. Int J Equity Health. 2023;22:248. 10.1186/s12939-023-02050-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Rehman M, Chapman L, Liu L, Calvert S, Sukhera J. Structural stigma within inpatient care for people who inject drugs: implications for harm reduction. Harm Reduct J. 2024;21:53. 10.1186/s12954-024-00971-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Mayampurath A, Hahs A, Caskey J, et al. 1497: Association between stigmatizing language occurrence in notes and outcomes in patients with sepsis. Crit Care Med. 2025;53. 10.1097/01.ccm.0001104652.97354.d8 [DOI] [Google Scholar]
- 6. Gupta P, Mayampurath A, Gruenloh T, et al. Prior emergency medical services utilization is a risk factor for in-hospital death among patients with substance misuse: a retrospective cohort study. BMC Emerg Med. 2024;24:110. 10.1186/s12873-024-01025-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Fleury MJ, Cao Z, Grenier G, Huỳnh C. Predictors of death from physical illness or accidental/intentional causes among patients with substance-related disorders. Can J Psychiatry Rev Can Psychiatr. 2023;68:163-177. 10.1177/07067437221136461 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Gruenloh T, Gupta P, Afshar AS, et al. Integrating multiple data sources to predict all-cause readmission or mortality in patients with substance misuse. PLOS Digit Health. 2025;4:e0001008. 10.1371/journal.pdig.0001008 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Afshar M, Oguss M, Callaci TA, et al. Creation of a data commons for substance misuse related health research through privacy-preserving patient record linkage between hospitals and state agencies. JAMIA Open. 2023;6:ooad092. 10.1093/jamiaopen/ooad092 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Inpatient Stays Involving Mental and Substance Use Disorders, 2016. #249 [Internet]. [accessed 2026 May 27]. https://hcup-us.ahrq.gov/reports/statbriefs/sb249-Mental-Substance-Use-Disorder-Hospital-Stays-2016.jsp [PubMed]
- 11. Collins GS, Moons KGM, Dhiman P, et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378. 10.1136/bmj-2023-078378 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Kind AJH, Buckingham WR. Making neighborhood-disadvantage metrics accessible—the neighborhood atlas. N Engl J Med. 2018;378:2456-2458. 10.1056/NEJMp1802313 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Elixhauser A, Steiner C, Harris DR, Coffey RM. Comorbidity measures for use with administrative data. Med Care. 1998;36:8-27. 10.1097/00005650-199801000-00004 [DOI] [PubMed] [Google Scholar]
- 14. Churpek MM, Yuen TC, Winslow C, et al. Multicenter development and validation of a risk stratification tool for ward patients. Am J Respir Crit Care Med. 2014;190:649-655. 10.1164/rccm.201406-1022OC [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Subbe CP, Kruger M, Rutherford P, Gemmel L. Validation of a modified early warning score in medical admissions. QJM Mon J Assoc Physicians. 2001;94:521-526. 10.1093/qjmed/94.10.521 [DOI] [PubMed] [Google Scholar]
- 16. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988;44:837-845. [PubMed] [Google Scholar]
- 17. Julian TH, Syeed R, Glascow N, Zis P. Alcohol-induced autonomic dysfunction: a systematic review. Clin Auton Res. 2020;30:29-41. 10.1007/s10286-019-00618-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data availability is subject to existing data use agreements. For further details, please contact the corresponding author.