Loading ...

Among hospitalized patients with cirrhosis, a machine learning (ML) model enhanced mortality prediction compared with traditional methods and was consistent across country income levels in a large global study.

“This highly inclusive, representative, and globally derived model has been externally validated,” Jasmohan Bajaj, MD, professor of medicine at Virginia Commonwealth University in Richmond, Virginia, told Medscape Medical News. “This gives us a crystal ball. It helps hospital teams, transplant centers, gastroenterology and intensive care unit services triage and prioritize patients more effectively.”

The study supporting the model, which Bajaj said “could be used at this stage,” was published online in Gastroenterology. The model is available for downloading at https://silveys.shinyapps.io/app_cleared/.

CLEARED Cohort Analyzed

Wide variations across the world regarding available resources, outpatient services, reasons for admission, and etiologies of cirrhosis can influence patient outcomes, according to Bajaj and colleagues. Therefore, they sought to use ML approaches to improve prognostication for all countries.

They analyzed admission-day data from the prospective Chronic Liver Disease Evolution And Registry for Events and Decompensation (CLEARED) consortium, which includes inpatients with cirrhosis enrolled from six continents. The analysis compared ML approaches with logistical regression to predict inpatient mortality.

The researchers performed internal validation (75/25 split) and subdivision using World-Bank income status: low/low-middle (L-LMIC), upper middle (UMIC), and high (HIC). They determined that the ML model with the best area-under-the-curve (AUC) would be externally validated in a US-Veteran cirrhosis inpatient population.

The CLEARED cohort included 7239 cirrhosis inpatients (mean age, 56 years; 64% men; median MELD-Na, 25) from 115 centers globally; 22.5% of centers belonged to LMICs, 41% to UMICs, and 34% to HICs.

A total of 808 patients (11.1%) died in the hospital.

Random-Forest analysis showed the best AUC (0.815) with high calibration. This was significantly better than parametric logistic regression (AUC, 0.774) and LASSO (AUC, 0.787) models.

Random-Forest also was better than logistic regression regardless of country income-level: HIC (AUC,0.806), UMIC (AUC, 0.867), and L-LMICs (AUC, 0.768).

Of the top 15 important variables selected from Random-Forest, admission for acute kidney injury, hepatic encephalopathy, high MELD-Na/white blood count, and not being in high income country were variables most predictive of mortality.

In contrast, higher albumin, hemoglobin, diuretic use on admission, viral etiology, and being in a high-income country were most protective.

The Random-Forest model was validated in 28,670 veterans (mean age, 67 years; 96% men; median MELD-Na,15), with an inpatient mortality of 4% (1158 patients).

The final Random-Forest model, using 48 of the 67 original covariates, attained a strong AUC of 0.859. A refit version using only the top 15 variables achieved a comparable AUC of 0.851.

Clinical Relevance

“Cirrhosis and resultant organ failures remain a dynamic and multidisciplinary problem,” Bajaj noted. “Machine learning techniques are one part of multi-faceted management strategy that is required in this population.”

If patients fall into the high-risk category, he said, “careful consultation with patients, families, and clinical teams is needed before providing information, including where this model was derived from. The results of these discussions could be instructive regarding decisions for transfer, more aggressive monitoring/ICU transfer, palliative care or transplant assessments.”

Meena B. Bansal, MD, system chief, Division of Liver Diseases, Mount Sinai Health System in New York City, called the tool “very promising.” However, she told Medscape Medical News, “it was validated on a VA [Veterans Affairs] cohort, which is a bit different than the cohort of patients seen at Mount Sinai. Therefore, validation in more academic tertiary care medical centers with high volume liver transplant would be helpful.” 

Furthermore, said Bansal, who was not involved in the study, “they excluded those that receiving a liver transplant, and while only a small number, this is an important limitation.”

Nevertheless, she added, “Artificial intelligence has great potential in predictive risk models and will likely be a tool that assists for risk stratification, clinical management, and hopefully improved clinical outcomes.”

This study was partly supported by a VA Merit review to Bajaj and the National Center for Advancing Translational Sciences, National Institutes of Health. No conflicts of interest were reported by any author.

Marilynn Larkin, MA, is an award-winning medical writer and editor whose work has appeared in numerous publications, including Medscape Medical News and its sister publication MDedge, The Lancet (where she was a contributing editor), and Reuters Health.