# Harnessing unsupervised machine learning with [18F]FDG PET/CT to develop a composite model for predicting overall survival in cervical cancer patients undergoing concurrent chemoradiotherapy

## TL;DR

This study uses PET/CT imaging and inflammation markers with a new clustering method to better predict survival in cervical cancer patients receiving chemoradiotherapy.

## Contribution

A novel unsupervised two-way clustering approach is introduced to combine radiomic features and inflammatory markers for improved survival prediction.

## Key findings

- The two-way clustering method outperformed traditional K-means in stratifying patients into distinct survival risk groups.
- Models using meta-features from two-way clustering achieved higher C-index values compared to PCA-based models.
- Combining clinical data, inflammatory markers, and radiomic features yielded the highest predictive accuracy with an AUC of 0.88.

## Abstract

This study sought to develop an advanced composite model to enhance the prognostic accuracy for cervical cancer patients undergoing concurrent chemoradiotherapy (CCRT). The model integrated imaging features from [18F]FDG PET/CT scans with inflammatory markers using a novel unsupervised two-way clustering approach.

In this retrospective study, 154 patients diagnosed with primary cervical cancer and treated with CCRT were evaluated using [18F]FDG PET/CT scans. A total of 1,702 radiomic features were extracted from the imaging data. These features underwent rigorous selection based on reproducibility and non-redundancy. The unsupervised two-way clustering method was then employed to simultaneously stratify patients and reduce the dimensionality of features, resulting in the generation of meta-features that were subsequently used to predict overall survival.

Kaplan-Meier survival analysis demonstrated that the two-way clustering method successfully stratified patients into distinct risk groups with significant survival differences (P<0.001), outperforming traditional K-means clustering. Predictive models constructed using meta-features derived from two-way clustering showed superior performance compared to those using principal component analysis (PCA), particularly when more than four features were included. The highest C-index values for the COX, COX_Lasso, and RSF models were observed with nine meta-features, yielding results of 0.691 ± 0.026, 0.634 ± 0.018, and 0.684 ± 0.020, respectively. In contrast, models based solely on clinical variables exhibited lower predictive performance, with C-index values of 0.645 ± 0.041, 0.567 ± 0.016, and 0.561 ± 0.033. The combination of clinical data, inflammatory markers, and radiomic features achieved the highest predictive accuracy, with a mean AUC of 0.88 ± 0.07.

Integrating radiomic data with inflammatory markers using unsupervised two-way clustering offered a robust approach for predicting survival outcomes in cervical cancer patients. This methodology presented a promising avenue for personalized patient management, potentially leading to more informed treatment decisions and improved outcomes.