Subsite-specific prediction of lymphatic spread in head-and-neck cancer using a mixture of hidden Markov models

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Head-and-neck squamous cell carcinomas (HNSCC) metastasize to neck lymph nodes. Radiation treatments include the elective irradiation of lymph node levels (LNLs) at risk of harboring occult (clinically undetected) metastases. We present a statistical model for personalized estimation of ipsilateral occult metastases risk given an individual patient’s clinical LNL involvement, T-stage, and tumor subsite. Each LNL is described by a binary random variable: healthy/metastatic. Lymphatic cancer progression is described by hidden Markov models (HMM) whose transition matrices contain the probabilities of tumor spread to and between LNLs. Specific primary tumor subsites (ICD-codes) are described by a mixture of HMMs. The model parameters are learned via the expectation–maximization (EM) algorithm from a multi-institutional dataset containing 2437 patients across 13 subsites. A mixture model with 4 HMMs identified components corresponding to the characteristic spread patterns associated with the anterior oral cavity, oropharynx, hypopharynx, and glottic larynx. The mixture coefficients describe a subsite’s similarity to each component and allow modeling gradual changes in LNL involvement for anatomically neighboring subsites with similar lymphatic spread. Palate tumors are described by mixtures between oropharynx and oral cavity, supraglottic larynx tumors as mixtures between glottic larynx and hypopharynx. The model predicts low risk of occult metastases in LNL III for cN0 oral cavity subsites, and low risk in LNL IV for oropharyngeal subsites without clinical LNL III involvement. This framework provides interpretable and individualized predictions of lymphatic spread in HNSCC. It may support the design of future clinical trials to investigate personalized volume-deescalated elective nodal irradiation strategies.