Association between immunohistochemical markers and tumor progression following resection of spinal chordomas: a...

Chordomas are slow-growing tumors derived from notochord remnants. Despite margin-negative excision and postoperative radiation therapy, spinal chordomas (SCs) often progress. The potential of immunohistochemical (IHC) markers, such as epithelial membrane antigen (EMA), combined with machine learning algorithms to predict long-term (≥ 12 months) postoperative tumor progression, has been understudied. The authors aimed to identify IHC markers using trained tree-based algorithms to predict long-term (≥ 12 months) postoperative tumor progression.

The authors reviewed the records of patients who underwent resection of SCs between January 2017 and June 2021 across the Mayo Clinic enterprise. Demographics, type of treatment, histopathology, and other relevant clinical factors were abstracted from each patient's record. Low tumor progression was defined as more than a 94.3-mm3 decrease in the tumor size at the latest radiographic follow-up. Decision trees and random forest classifiers were trained and tested to predict the long-term volumetric progression after an 80/20 data split.

Sixty-two patients diagnosed with and surgically treated for SC were identified, of whom 31 were found to have a more advanced tumor progression based on the tumor volume change cutoff of 94.3 mm3. The mean age was 54.3 ± 13.8 years, and most patients were male (62.9%) and White (98.4%). The most common treatment modality was subtotal resection with radiation therapy (35.5%), with proton beam therapy being the most common (71%). Most SCs were sacrococcygeal (41.9%), followed by cervical (32.3%). EMA-positive SCs had a postoperative progression risk of 67%. Pancytokeratin-positive SCs had a progression rate of 67%; however, patients with S100 protein-positive SCs had a 54% risk of progression. The accuracy of this model in predicting the progression of unseen test data was 66%. Pancytokeratin (mean minimal depth = 1.57), EMA (mean minimal depth = 1.58), cytokeratin A1/A3 (mean minimal depth = 1.59), and S100 protein (mean minimal depth = 1.6) predicted the long-term volumetric progression. Multiway variable importance plots show the relative importance of the top 10 variables based on three measures of varying significance and their predictive role.

These IHC variables with tree-based machine learning tools successfully demonstrate a high capacity to identify a patient's tumor progression pattern with an accuracy of 66%. Pancytokeratin, EMA, cytokeratin A1/A3, and S100 protein were the IHC drivers of a low tumor progression. This shows the power of machine learning algorithms in analyzing and predicting outcomes of rare conditions in a small sample size.