Purpose: Evaluate the performance of machine learning (ML) models to predict the time until a head and neck cancer patient will recur based on information from CT and PET images.
Methods: 201 head & neck patients made available by the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) were used for this study. For each patient, 110 GTV features were collected from PET and CT images. Their GTVs were annotated according to procedure outlined by the MICCAI. Logistic regression was used to predict the probability that a patient recurs. This probability, along with 220 features (110 PET and 110 CT) were used as input to a lasso regression and COX proportional hazards model to predict time to recurrence (TTR). The concordance index was the metric used to compare different models.
Results: Compared to lasso regression, the COX proportional hazards model performed significantly better with a concordance index of 0.66 versus 0.56 with a corresponding Chi-squared value of 7.517 (P-value = 0.00611). Including the probability of recurrence predicted by logistic regression, and accounting for censorship did not improve TTR prediction. However, including image information from both PET and CT images, as opposed to just one modality, did improve model performance increasing the concordance index from 0.57 and 0.58 for PET and CT data only respectively, to 0.66 with a corresponding Chi-squared value of 5.795 (P-value = 0.016) and 10.357 (P-value = .00129).
Conclusion: This work emphasizes the implementation of ML models to predict important treatment outcomes. Specifically, the COX model is a promising method to predict when a patient will recur. Model improvement can be further investigated with the inclusion of clinical and biopsy data. This work adds to the growing acknowledgement that ML approaches have the potential to aid the prognosis of cancer patients.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by a research grant from Varian Medical Systems (Palo Alto, CA, USA), UL1TR001433 from the NCATS/NIH and R21EB030209 from the NIBIB/NIH. The content is solely the responsibility of the author and does not necessarily represent the official views of the NIH.