Systematic review on risk prediction models prior to non-cardiac surgeries for the assessment of post-operative mortality and morbidity

National and international clinical guidelines recommend the use of risk prediction models to assess postoperative risk of major adverse cardiac events (MACE) - including death, stroke, and myocardial infarction (MI) - in adults undergoing non-cardiac surgery [1-4]. Revised Cardiac Risk Index (RCRI) is recommended as a first step in preoperative risk assessment [5], but other models are feasible, too. Due to the large number of available risk prediction models [2] with different predictive properties, selecting an appropriate model for individual patients can be challenging for clinicians [5-8]. 

The presented systematic review aimed to provide an overview over the existing models including their specific characteristics for the prediction of MACE as primary outcome and rehospitalization, acute kidney injury, or infection as secondary outcomes. 

A systematic review was performed following the “Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA)” statement [9]. Eligible studies reported on risk prediction models and corresponding performance measures for adult patients undergoing inpatient non-cardiac surgery. The prediction needed to be performed prior to surgery. Outcomes needed to be assessed during surgery or within 30 days afterwards. The review was registered in PROSPERO (CRD42024546521). Four databases were searched for evidence published until May 2024. Abstract and full-text screening as well as data extraction and risk of bias assessment was performed by two independent reviewers following standardized reporting guidelines [10, 11], conflicts were resolved by discussion.

After database search, 930 unique records were identified and assessed for eligibility;  89 articles were included in data extraction. Preliminary results showed that 161 different risk prediction models were reported. The most commonly reported model was the RCRI (n=29), followed by American Society of Anesthesiologists (ASA) physical status (n=9). The most reported biomarker used for risk prediction was NT-proBNP (n=11). Most publications (of new models) did not include any validation for the described models. MACE as a composite outcome was predicted in 20 articles. Due to heterogenous definitions of MACE, components were analyzed separately. Death was the most frequently reported single outcome (n=144), followed by stroke (n=16) and MI (n=11). Acute kidney injury was the most frequently reported secondary outcome (n=13). Overall, 205 single and 162 combined outcomes were reported. Model discrimination varied greatly across articles and subgroups of patients, i.e. RCRI c-statistic 0.457 to 0.942. Only 13.5% of the included articles had low risk of bias according to the PROBAST tool; most studies had high risk of bias (56.2%). Reporting of model performance focused primarily on discrimination and only a minority of publications reported model calibration. The number of events per variable was too small (<20) for reliable results in several articles (n=33) according to the PROBAST guideline.

Our results showed a large number of existing risk prediction models and differences in model performance depending on the patient cohort. It further confirms the shortcomings in model performance reporting. We found substantial heterogeneity in models, which highlights the need for standardization, improved methodological quality and validation of existing models before clinical implementation.