Modelling occult lymph node metastases in HNSCC patients with a trinary state hidden Markov model.

Objective.Head-and-neck squamous cell carcinoma frequently metastasize through lymphatic system. Occult metastases are challenging for designing radiotherapy treatment volumes. Literature values indicate that around 20% of lymph node metastases are clinically undetected. However, recent data suggest that this value is only representative for level II whereas the rate of occult metastases is substantially higher in levels III and IV.Approach.We propose a trinary-state Hidden Markov Model to describe ipsilateral lymphatic tumor progression over time. Each lymph node level (LNL) can be in one of three states: healthy, microscopically (pathologically) involved, or macroscopically (clinically) involved. In each time step, a healthy LNL may become microscopically involved due to spread from the primary tumor or an involved upstream LNL. In addition, a microscopically involved LNL may transition to macroscopically involved. The probability of occult metastases is obtained as the conditional probability of being in the microscopically involved state given the individual patient's macroscopic involvement. Model parameters are learned from a dataset of 550 patients, including 263 with both pathological and clinical LNL involvement reported.Main results.For oropharyngeal SCC, the model estimates an occult metastases probability below 5% in LNL IV unless LNL III is clinically positive, suggesting potential for reducing elective clinical target volumes. The model's estimated rate of clinically undetected metastases is 82%, 41%, and 34% for LNL II, III, and IV, respectively, which agrees with the data.Significance.The proposed trinary-state HMM represents a methodological extension to a previously published binary-state HMM. The binary HMM distinguished the microscopic and macroscopic involvement via the concept of sensitivity, which may underestimate the risk of occult metastases. The trinary HMM addresses this problem and represents a more natural description of tumor progression.