Personalized prediction of postoperative complication and survival among Colorectal Liver Metastases Patients Receiving Simultaneous Resection using machine learning approaches: A multi-center study.

BACKGROUND

To predict clinical important outcomes for colorectal liver metastases (CRLM) patients receiving colorectal resection with simultaneous liver resection by integrating demographic, clinical, laboratory, and genetic data.

METHODS

Random forest (RF) models were developed to predict postoperative complications and major complications (binary outcomes), as well as progression-free survival (PFS) and overall survival (OS) (time-to-event outcomes) of the CRLM patients based on data from two hospitals. The models were validated on an external dataset from an independent hospital. The clinical utility of the models was assessed via decision curve analyses (DCA).

RESULTS

There were 1067 patients included in survival prediction analyses and 1070 patients included in postoperative complication prediction analyses. The RF models provided an assessment of the model contributions of features for outcomes and suggested KRAS, BRAF, and MMR status were salient for the PFS or OS predictions. RF model of PFS showed that the Brier scores at 1-, 3-, and 5-year PFS were 0.213, 0.202 and 0.188; and the AUCs of 1-, 3- and 5-year PFS were 0.702, 0.720 and 0.743. RF model of OS revealed that Brier scores of 1-,3-, and 5-year OS were 0.040, 0.183 and 0.211; and the AUCs of 1-, 3- and 5-year OS were 0.737, 0.706 and 0.719. RF model for postoperative complication resulted in an AUC of 0.716 and a Brier score of 0.196. DCA curves clearly demonstrated that the RF models for these outcomes exhibited a superior net benefit across a wide range of threshold probabilities, signifying their favorable clinical utility. The RF models consistently exhibited robust performance in both internal cross-validation and external validation. The individualized risk profile predicted by the models closely aligned with the actual survival outcomes observed for the patients. A web-based tool (https://kanli.shinyapps.io/CRLMRF/) was provided to demonstrate the practical use of the prediction models for new patients in the clinical setting.

CONCLUSION

The predictive models and a web-based tool for personalized prediction demonstrated a moderate predictive performance and favorable clinical utilities on several key clinical outcomes of CRLM patients receiving simultaneous resection, which could facilitate the clinical decision-making and inform future interventions for CRLM patients.