New prediction models may help clinicians identify which patients are at highest risk for death or drug overdose if they seek a "before medically advised" (BMA) hospital discharge.

In a population-based retrospective cohort study using data from British Columbia, factors associated with higher risk for death after BMA discharge included a Charlson Comorbidity Index of 2 or higher, cancer, and heart disease. Factors associated with higher risk for overdose following BMA discharge were homelessness, receipt of social income assistance, opioid use disorder, nonalcohol substance use disorder, overdose in the past year, and discharge from a surgical service. 

"Although our study findings indicated that adverse outcomes after BMA discharge may be predictable, further research is required to determine whether these outcomes are preventable and by what means," wrote Hiten Naik, MD, an internal medicine physician and research fellow at the University of British Columbia in Vancouver, and co-authors.

The data were published November 11 in the Canadian Medical Association Journal.

Two Models 

To develop means for estimating patients' risk for adverse outcomes after BMA discharge, the researchers examined the BC Provincial Overdose Cohort's Reference Cohort (ODCr), which includes data from a random sample representing 20% of all British Columbia residents.

British Columbia provides publicly funded health insurance, which entails free access to emergency care, hospital services, and primary care. This program aids the collection of patient-level data on medical care. The ODCr includes data on treatment for overdoses, community mental health service use, substance use history, social income assistance, homelessness, and date of death. 

Using this robust database, the researchers created two models. Model A was intended to predict death from any cause in the first 30 days after BMA discharge. It was based on Cohort A, which included all nonelective, nonobstetrical adult hospital admissions in the database that ended in BMA discharge between 2015 and 2019. Cohort A included 6440 hospitalizations with BMA discharge. 

Model B was intended to predict fatal and nonfatal illicit drug overdose in the first 30 days after BMA discharge among patients with a history of substance use. The model was based in Cohort B, which included all hospital admissions from cohort A for patients with evidence of prior substance use (i.e., an overdose in the past year, any substance use disorder diagnosis in the past 5 years, or a prescription fill for opioid agonist therapy in the past 5 years). Cohort B included 4466 hospitalizations with BMA discharge. 

In Cohort A, death after BMA discharge "was less common than clinicians might expect (i.e., 1 death within 30 days for every 63 BMA discharges)," the authors wrote. "We found that multimorbidity, heart disease, and cancer were strong predictors of death from any cause within 30 days of BMA discharge." 

In Cohort B, the researchers found about one illicit drug overdose within 30 days for every 19 BMA discharges, "suggesting this period is a critical but largely unexplored opportunity for overdose prevention," they wrote.

Based on this research, the investigators developed risk calculators intended to predict individual patients' risk for death and illicit drug overdose after BMA discharge. With further research, these models could help hospitals automate their responses to higher-risk BMA discharges, they wrote.

Thoughtful Consideration Required 

The deployment of risk models could make it easier for physicians to have difficult conversations with patients who are intent on leaving the hospital early. "I've always had moral distress and internal questions about whether I'm doing the right thing and what responsibilities I have with these patients," Naik told Medscape News Canada.

Many patients who initiate BMA discharges have a history of substance abuse, and the stigma associated with this history influences clinicians' care for these patients, Naik said. But clinicians need to approach conversations about BMS discharges with the same care that they use to decide on medicines to help patients understand the risks that these prescriptions entail. 

"Sometimes we just don't have a habit of going through that thought process" with patients who reject physicians' advice and leave hospitals early, said Naik. "Sometimes you may think, 'It's not our problem anymore to resolve.' Through the publication of this paper and the posting of the online calculators, we are highlighting the fact that it's an important clinical issue that requires thoughtful consideration and communication." 

Among other limitations of their study, the researchers acknowledged that they "did not compare model performance to clinical gestalt, which can sometimes perform as well as, or better than, some clinical prediction models."

The investigators are seeking additional funding to study the effectiveness of these models with more recent data. It also would be important for other researchers to apply the models to other datasets. "These prediction models require external validation before widespread clinical use," Naik and colleagues wrote.

External Validation Essential 

Prediction models that have not been validated pose challenges for clinicians and researchers, Farid Foroutan, PhD, associate scientist at the Toronto General Hospital Research Institute, told Medscape News Canada. "There's a lot of development, and very few [models] get externally validated. That's a pet peeve of ours about the research landscape."

Foroutan, who did not participate in the research, is also an associate scientist at the Ted Rogers Computation Program at Toronto's University Health Network. 

To externally validate the models proposed by Naik and co-authors, researchers could investigate outcomes among patients in the ODCr during a different period. Researchers in other parts of Canada also could apply the model to their own independent patient cohorts and evaluate how well it predicts death 30 days after BMA discharge, Foroutan said.

In many cases, randomized clinical trials would be the ultimate way to see if a prediction model aids medical care, he said. "Physicians are learning a lot about their patients and have an easier time making predictions about them," especially in cases where a physician manages a patient's chronic condition such as heart failure. "If you want physicians to use a model, then that model has to be much better than the physician's intuition to begin with."

The study was supported by the Canadian Institutes of Health Research, the Vancouver Coastal Health Research Institute, and the University of British Columbia Division of General Internal Medicine. Naik was supported by a CAN-TAP-TALENT and Michael Smith Health Research BC Postdoctoral Fellowship Award, and a CIHR Fellowship Award. Foroutan reported no relevant financial relationships.

Kerry Dooley Young is a freelance journalist based in Washington, DC. She has reported on medical research and health care policy for more than 20 years.