Hidden patterns, different outcomes: un-supervised machine-learning clustering in oligometastatic gynecologic cancer

To stratify the treatment outcomes of patients with oligometastatic gynecologic cancer receiving stereotactic body radiation therapy using an un-supervised clustering machine-learning method.

This multi-centric study was based on a cohort of 172 patients receiving curative-intent stereotactic body radiation therapy for oligometastatic uterine tumors, yielding a total of 268 lesions. The following clinical and dosimetric variables were collected: age, number of lesions per patient, type of lesion (lymph nodes vs parenchyma), lesion burden (number of treated lesions per patient), treatment site of lesion, number of fractions, total dose, biologically effective dose, and planning target volume. An un-supervised clustering method based on the K-means algorithm was used to identify clusters of lesions. The groups of lesions were compared in terms of local control, distant-metastases-free survival, and overall survival.

The optimal number of clusters was found to be equal to 3. The analysis of variance indicated that the variables contributing the most to the separation of the clusters were the planning target volume, the biologically effective dose, and the type of lesion. Significant differences were found between the 3 groups of lesions in terms of local control (p =.002). At 2 years, local control was 84.6%, 74.7%, 47.5% for the 3 clusters that were "a posteriori" named as high-control, medium-control, and low-control, respectively. Distant-metastases-free survival was also found to be significantly different (p =.04) at 2 years, with values of 27.0%, 22.2%, 48.0% for the high-control, medium-control, and low-control, respectively. No differences were found for the overall survival (p =.22).

In this study, un-supervised machine-learning partitioned oligometastatic lesions into 3 clusters associated with different treatment responses. A prospective validation is needed for prediction purposes.