Accurate long‐term prediction of death for patients with cirrhosis

Abstract

Background and Aims: Cirrhosis is a major cause of death and is associated with extensive health care use. Patients with cirrhosis have complex treatment choices due to risks of morbidity and mortality. To optimally counsel and treat patients with cirrhosis requires tools to predict their longer-term liver-related survival. We sought to develop and validate a risk score to predict longer-term survival of patients with cirrhosis. Approach and Results:We conducted a retrospective cohort study of adults with cirrhosis with no major life-limiting comorbidities. Adults with cirrhosis within the Veterans Health Administration were used for model training and internal validation, and external validation used the OneFlorida Clinical Research Consortium. We used four model-building approaches including variables predictive of cirrhosis-related mortality, focused on discrimination at key time points (1, 3, 5, and 10 years). Among 30,263 patients with cirrhosis ≤75 years old without major life-limiting comorbidities and complete laboratory data during the baseline period, the boosted survival tree models had the highest discrimination, with 1-year, 3-year, 5-year, and 10-year survival rates of 0.77, 0.81, 0.84, and 0.88, respectively. The 1-year, 3-year, and 5-year discrimination was nearly identical in external validation. Secondary analyses with imputation of missing data and subgroups by etiology of liver disease had similar results to the primary model. Conclusions:We developed and validated (internally and externally) a risk score to predict longer-term survival of patients with cirrhosis. This score would transform management of patients with cirrhosis in terms of referral to specialty care and treatment decision-making for non-liver-related care.

Prediction of long-term survival among patients with cirrhosis using time-varying models

“…This is a statewide database of clinical (eg, administrative codes) and laboratory data from 11 health care systems and affiliated practices (inpatient and outpatient) that provide health care to >40% of Floridians, which, over time, has represented >15 million unique patients. 5 , 9 – 11 The study period for patient identification and follow-up was from January 1, 2015, to December 31, 2019.…”

“…Over the last several years, several risk scores focused on predicting survival over relevant intermediate and longer term periods and have been developed with excellent prediction. 3 – 5 The first score was developed in the Veterans Health Administration (VHA) but included patients with life-limiting conditions in their cohort and a composite comorbidity score in their model, thereby limiting it as a tool to predict liver-specific mortality. 4 , 6 A second score used the baseline MELD score and 2-dimensional shear wave elastography but performed best by categorizing patients as “good,” “intermediate,” or “poor” prognosis over a 2-year horizon.…”

“… 7 A more recent score developed in the VHA and externally validates in the OneFlorida Clinical Research Consortium overcame the limitations of these other scores by excluding patients with life-limiting comorbidities and focusing on clinical and laboratory data obtained in routine clinical practice. 5 However, all 3 of these scores have a key limitation as it relates to clinical care—they rely only on baseline clinical and laboratory data, rather than time-updating information that reflects the natural history and different clinical trajectories of a given patient. Therefore, we sought to develop a clinically relevant risk score to predict intermediate and longer term survival of patients with cirrhosis using time-updating clinical (ie, complications of portal hypertension) and laboratory data.…”

The computer will see you now: Prediction of long‐term survival in patients with cirrhosis

“…The building blocks of AI in cirrhosis have been laid by recent investigations. [1,4,6] It is now imperative for us to proceed in a thoughtful and nuanced manner to complement (rather than replace) our clinical assessment with AI predictive tools to achieve the eventual goal of personalized medicine.…”

“…[1][2][3][4][5] In this issue of Hepatology, Goldberg and colleagues developed a prediction model of long-term survival (long-term cirrhosis survival [LTCS]) utilizing machine-learning-boosted survival trees techniques in a well-characterized cohort within the Veterans Health Administration followed by external validation using the OneFlorida Clinical Research Consortium. [6] Patients with hepatocellular carcinoma, nonliverlimiting comorbidities, and absolute contraindications for liver transplant were excluded to better characterize cirrhosis-related mortality in potential transplant candidates.…”