Personalized prognostic model for colorectal cancer in the era of precision medicine: a dynamic approach based on real-world data.

BACKGROUND

Predicting individual prognosis is required for patients with colorectal cancer in the era of precision medicine. However, this may be challenging for the conventional survival analysis such as the Cox proportional hazards model. This study aims to develop a personalized prognostic prediction that incorporates longitudinal data to improve predictions for colorectal cancer patients.

METHODS

Patients with advanced or recurrent colorectal cancer, who received treatment at Kyoto University Hospital between April 2015 and December 2021, were retrospectively analyzed. The Joint model is one of the dynamic prediction models. Using longitudinal clinical data, a carcinoembryonic antigen (CEA) prediction equation was developed for each patient. Additionally, a personalized prognostic prediction model was created using the Joint model. The prediction accuracy of the Joint model was compared with one of the Cox proportional hazards model.

RESULTS

Among the 1010 patients, 614 patients were enrolled. The median frequency of tumor marker measurement (per patient) was 20 times (range: 3-117 times). CEA values could be predicted accurately and the Pearson's correlation coefficient between measured CEA and predicted CEA was 0.931. In the Joint model, the significant prognostic factors were baseline age (HR, 1.039; 95% CI, 1.025-1.054), poor-differentiated tumor (HR, 2.600; 95% CI 1.446-4.675) and log (predicted CEA) (HR, 1.551; 95% CI 1.488-1.617). The areas under the curve at 2, 3, 4, and 5 were significantly higher for the Joint model than for the Cox proportional hazards model, respectively.

CONCLUSION

The Joint model may accurately predict personalized prognosis that reflects changes in longitudinal tumor marker values.