Assessing overall tumor burden on the basis of tumor number and size may assist in prognostic stratification of patients after resection of colorectal liver metastases (CRLM). We sought to define the prognostic accuracy of tumor burden by using machine learning (ML) algorithms compared with other commonly used prognostic scoring systems.Patients who underwent hepatectomy for CRLM between 2001 and 2018 were identified from a multi-institutional database and split into training and validation cohorts. ML was used to define tumor burden (ML-TB) based on CRLM tumor number and size thresholds associated with 5-year overall survival. Prognostic ability of ML-TB was compared with the Fong and Genetic and Morphological Evaluation scores using Cohen's d.Among 1,344 patients who underwent resection of CRLM, median tumor number (2, interquartile range 1 to 3) and size (3 cm, interquartile range 2.0 to 5.0) were comparable in the training (n = 672) vs validation (n = 672) cohorts; patient age (training 60.8 vs validation 61.0) and preoperative CEA (training 10.2 ng/mL vs validation 8.3 ng/mL) was also similar (p > 0.05). ML empirically derived optimal cutoff thresholds for number of lesions (3) and size of the largest lesion (1.3 cm) in the training cohort, which were then used to categorize patients in the validation cohort into 3 prognostic groups. Patients with low, average, or high ML-TB had markedly different 5-year overall survival (51.6%, 40.9%, and 23.1%, respectively; p < 0.001). ML-TB was more effective at stratifying patients relative to 5-year overall survival (low vs high ML-TB, d = 2.73) vs the Fong clinical (d = 1.61) or Genetic and Morphological Evaluation (d = 0.84) scores.Using a large international cohort, ML was able to stratify patients into 3 distinct prognostic categories based on overall tumor burden. ML-TB was noted to be superior to other CRLM prognostic scoring systems.
View details for DOI 10.1097/XCS.0000000000000094
View details for PubMedID 35290269