Development of an open-source tool for risk assessment in pulmonary endarterectomy.

Background: Risk prediction tools are routinely utilised in cardiothoracic surgery but have not been developed for pulmonary endarterectomy (PEA). There is no data on whether patients undergoing PEA may benefit from a tailored risk modelling approach. We develop and validate a clinically-usable tool to predict PEA 90-day mortality (90 DM) with the secondary aim of informing factors that may influence five-year mortality (5 YM) and improvement in patient-reported outcomes (PROchange) using common clinical assessment parameters. Derived model predictions were compared to those of the currently most widely implemented cardiothoracic surgery risk tool, EuroSCORE II.

Methods: Consecutive patients undergoing PEA for chronic thromboembolic pulmonary hypertension (CTEPH) between 2007 and 2018 (n=1334) were included in a discovery dataset. Outcome predictors included an intentionally broad array of variables, incorporating demographic, functional and physiological measures. Three statistical models (linear regression, penalised linear regression and random forest) were considered per outcome, each calibrated, fitted and assessed using cross-validation, ensuring internal consistency. The best predictive models were incorporated into an open-source PEA risk tool and validated using a separate prospective PEA cohort from 2019 to 2021 (n=443) at the same institution.

Results: Random forest models had the greatest predictive accuracy for all three outcomes. Novel risk models had acceptable discriminatory ability for outcome 90 DM (AUROC 0.82) outperforming that of EuroSCORE II (AUROC 0.65). CTEPH related factors were important for outcome 90 DM but 5 YM was driven by non-CTEPH factors, dominated by generic cardiovascular risk. We were unable to accurately predict a positive improvement in PRO status (AUROC 0.47).

Conclusions: Operative mortality from PEA can be predicted pre-operatively to a potentially clinically useful degree. Our validated models enable individualised risk stratification at clinician point-of-care to better inform shared decision making.