Table 2. Key characteristics of included studies.

Lead author (year)	Setting	Aim	Principle methods	Data and sample	Key Results
Einav (2018) 33	US: random sample of Medicare fee-for- service beneficiaries in 2008	To analyse healthcare spending by predicted 12-mortality, i.e. can high end-of-life care costs be identified ex ante	Ensemble of RF, gradient boosting and LASSO	Administrative data: demographics, ICD codes, chronic conditions, prior utilization for a baseline sample of 5,631,168 Trajectories of health care use and diagnosis in the prior 12-month period were included	ML model attributed a higher risk score to those who died within one year than those who did not in 87% of cases End-of-life spending is high but deaths do not account heavily for high spending in Medicare overall Focusing on end-of-life spending is not a useful way to identify inappropriate treatment choices
Makar (2015) 34	US: Medicare fee-for- service beneficiaries in 2010	To quantify six-month mortality risk in four disease cohorts: cancer, COPD, CHF, dementia.	Six ML approaches and logistic regression were used in each cohort, of which RF models performed best in primary analysis.	Administrative data: demographics, ICD codes, chronic conditions, functional status, durable medical equipment, prior utilization for 20,000 randomly selected subjects in each disease cohort Traditional baseline characteristics were augmented with values in prior 12-month period, thus capturing disease progression, functional decline, etc.	ML model attributed a higher risk score to those who died within six months than those who did not in 82% of cases Augmented variables key to predictive power; models using only traditional variables were less accurate
Sahni (2018) 35	Minnesota, US: Six-hospital network (one large tertiary care centre; five community hospitals), 2012–2016	To quantify 1-year mortality risk in a cohort of clinically diverse hospitalized patients.	RF models and logistic regression were applied separately and performance compared.	Electronic medical record data, including vital signs, blood count, metabolic panel, demographics and ICD codes for 59,848 patients	ML model attributed a higher risk score to those who died within one year than those who did not in 86% of cases RF model outperforms logistic regression Demographic and lab data key to predictive power; models using ICD codes alone are less reliable

US: United States; RF: random forest; ML: machine learning; COPD: chronic obstructive pulmonary disease; CHF: congestive heart failure; ICD: international classification of disease; SEER: Surveillance, Epidemiology and End Results.