Predicting imminent disease progression in advanced colorectal cancer by a machine-learning algorithm.

Abstract

645 Background: In advanced cancers, predicting disease progression just before its clinical manifestation enables an earlier switch to the next treatment line, preventing deterioration in the patient's state and potentially improving survival. Yet, given the ambiguity of current tumor markers in alerting to progression, physicians are unable to forecast this key event. We developed a diagnostic algorithm for announcing an approaching disease progression in late-stage colorectal cancer (CRC) patients by processing continuous carcinoembryonic antigen (CEA) input. Methods: Longitudinally measured CEA data of advanced CRC patients treated by standard 1st line chemotherapies, collected from 2 clinical trials (projectdatasphere.org), served for algorithm development by machine-learning and training assisted by receiver-operating-characteristic (ROC) analysis and correlation tests. Performance was validated by cross-validation techniques. Results: CEA and response evaluations of 489 CRC patients (median follow-up time: 168 days) were processed by the algorithm, predicting disease progression with 57% sensitivity (100/175 progression events) and 88% specificity (21/175 false positives). Positive and negative predictive values, accuracy and Cohen’s kappa were 64%, 84%, 79% and 0.46, respectively. The algorithm’s predictive power was superior to that of standard statistical analyses of these CEA data (e.g., ROC). Conclusions: Our study offers a new approach to using tumor markers as prognosticators. The algorithm-amplified ability of CEA to predict progression in CRC complements our recent findings in lung cancer, where integration of CEA and 4 other markers provided 66% sensitivity in predicting progression, surpassing the low capacity of each separate marker. Conceivably, future algorithm-integration of multiple markers in CRC may also exceed the limited signal of a single marker. Clinical use of our algorithm, amplifying weak marker signals of imminent progression, should allow physicians to reliably harness tumor markers for improving treatment and potentially extending survival in cancer patients.