Leveraging automated machine learning to predict colon cancer prognosis from clinical features and risk groups: a retrospective cohort study

Abstract

Background

Predicting colon cancer recurrence is crucial for determining the need for adjuvant therapy after curative resection. However, clinical decisions often rely on limited features, even when a large amount of data is available.

Methods

We assessed the clinical utility of automated machine learning (AutoML) models to predict the prognosis of colon cancer patients from a tertiary hospital using clinical features, pathologic characteristics, and blood markers. We also compared these AutoML models to manually trained and tuned models and evaluated survival predictions.

Results

We found comparable performance between linear and ensemble models, and the predicted prognosis was significantly associated with overall survival and disease-free survival outcomes. Interpretable machine learning models identified T and N staging as important features and highlighted the prognostic immune and nutritional index (PINI) as a meaningful biomarker. The XGBoost model predicted prognosis with an AUC of 0.798 in an independent test set from a different hospital, demonstrating the model's interoperability. Additionally, the model was able to distinguish stage IIA patients that would benefit from adjuvant chemotherapy, a complex and difficult decision for clinicians. We also showed that simplified models generally maintained predictive accuracy, and that the automated approach was equally predictive as manually curated models.

Conclusion

With extensive validation through multiple test sets and internal cross-validation, this work underscores the potential of AutoML in identifying survival-related signatures in colon cancer from routinely collected data, providing clinicians with valuable insights for personalized treatment strategies.